slides-pt-br.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "# 1 + 1 = 1 ou Pareamento de Registros com Python\n",
    "Flávio Juvenal  \n",
    "[@flaviojuvenal](https://twitter.com/flaviojuvenal)  \n",
    "\n",
    "<p>\n",
    "    Sócio e Desenvolvedor na <a href=\"http://www.vinta.com.br?dedupe-slides=1\"><img class=\"vinta_logo\" style=\"display: inline-block; margin-left: 0;\" src=\"vinta.png\" width=300></a>\n",
    "</p>\n",
    "\n",
    "Outras palestras da Vinta em: http://bit.ly/pybr-talks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "import warnings; warnings.simplefilter('ignore')\n",
    "import logging; logging.disable(level=logging.INFO)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Use isto para setar GOOGLE_API_KEY e ativar o geocoder\n",
    "# Pegue uma em: https://cloud.google.com/maps-platform/#get-started\n",
    "# Depois crie um arquivo .env com GOOGLE_API_KEY=valor-da-key\n",
    "# E por fim rode as linhas comentadas abaixo:\n",
    "\n",
    "from dotenv import load_dotenv\n",
    "load_dotenv(verbose=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<p style=\"font-size: 2em; text-weight: bold\">\n",
    "    Slides disponíveis em: <a href=\"http://bit.ly/pybr-duplica\">http://bit.ly/pybr-duplica</a>\n",
    "</p>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Registros reais são uma bagunça!\n",
    "Provavelmente você já lidou com registros assim antes:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "scrolled": false,
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "\n",
    "data = [\n",
    "    (\"Reteteu\",\"R. Prof. Otávio de Freitas\",\"Recife - PE\"),\n",
    "    (\"Reteteu - Comida Honesta\",\"R. Prof. Otavio de Freitas, 256\",\"Recife (PE)\"),\n",
    "    (\"Reteteu (Restaurante)\",\"Rua Otávio de Freitas\",\"Encruzilhada (Recife)\"),\n",
    "    (\"Tio Pepe (Restaurante)\", \"R. Alm. Tamandaré, 170\", \"Boa Viagem (Recife)\"),\n",
    "    (\"Tio Pepe\", \"Rua Almirate Tamamdare\", \"Recife\"),\n",
    "]\n",
    "\n",
    "df = pd.DataFrame(data, columns=['restaurant', 'address', 'city'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>restaurant</th>\n",
       "      <th>address</th>\n",
       "      <th>city</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>Reteteu</td>\n",
       "      <td>R. Prof. Otávio de Freitas</td>\n",
       "      <td>Recife - PE</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Reteteu - Comida Honesta</td>\n",
       "      <td>R. Prof. Otavio de Freitas, 256</td>\n",
       "      <td>Recife (PE)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Reteteu (Restaurante)</td>\n",
       "      <td>Rua Otávio de Freitas</td>\n",
       "      <td>Encruzilhada (Recife)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Tio Pepe (Restaurante)</td>\n",
       "      <td>R. Alm. Tamandaré, 170</td>\n",
       "      <td>Boa Viagem (Recife)</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Tio Pepe</td>\n",
       "      <td>Rua Almirate Tamamdare</td>\n",
       "      <td>Recife</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                 restaurant                          address  \\\n",
       "0                   Reteteu       R. Prof. Otávio de Freitas   \n",
       "1  Reteteu - Comida Honesta  R. Prof. Otavio de Freitas, 256   \n",
       "2     Reteteu (Restaurante)            Rua Otávio de Freitas   \n",
       "3    Tio Pepe (Restaurante)           R. Alm. Tamandaré, 170   \n",
       "4                  Tio Pepe           Rua Almirate Tamamdare   \n",
       "\n",
       "                    city  \n",
       "0            Recife - PE  \n",
       "1            Recife (PE)  \n",
       "2  Encruzilhada (Recife)  \n",
       "3    Boa Viagem (Recife)  \n",
       "4                 Recife  "
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "A solução se chama **Pareamento de Registros** ([Record Linkage](https://en.wikipedia.org/wiki/Record_linkage)). Quando aplicado em um único dataset, se chama **Deduplicação de Registros** (Record Deduplication).\n",
    "\n",
    "Consiste em comparar e unir registros de maneira **fuzzy**, utilizando campos imprecisos como nomes, endereços, telefones, datas, etc."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Não confundir com Deduplicação de _Dados_, que é associado a compressão e armazenamento."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Comparando strings de forma fuzzy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "import jellyfish"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.7959224985540775"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "jellyfish.jaro_winkler(\"R. Prof. Otávio de Freitas\", \"Rua Otávio de Freitas\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.473970473970474"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "jellyfish.jaro_winkler(\"R. Prof. Otávio de Freitas\", \"R. Alm. Tamandaré, 170\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Comparando endereços de forma fuzzy"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Geocodificar endereços, i.e., convertê-los para **latitude/longitude** é bastante útil, pois possibilita:\n",
    "- limpar variações irrelevantes (prof. vs. professor)\n",
    "- calcular distância geométrica (mesma rua: mais perto, mais similar)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [],
   "source": [
    "import requests\n",
    "import geocoder\n",
    "\n",
    "full_addresses = [\n",
    "    \"R. Alm. Tamandaré, 170, Boa Viagem (Recife)\",\n",
    "    \"Rua Almirate Tamamdare, Recife\",\n",
    "]\n",
    "\n",
    "full_addresses_latlng = []\n",
    "with requests.Session() as session:\n",
    "    for a in full_addresses:\n",
    "        a_geocoded = geocoder.google(a, session=session)\n",
    "        full_addresses_latlng.append((a_geocoded.latlng, a_geocoded.postal))\n",
    "\n",
    "address_latlng = list(zip(full_addresses, full_addresses_latlng))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[('R. Alm. Tamandaré, 170, Boa Viagem (Recife)',\n",
      "  ([-8.1340452, -34.9083392], '51030-090')),\n",
      " ('Rua Almirate Tamamdare, Recife', ([-8.1338926, -34.9082632], '51030'))]\n"
     ]
    }
   ],
   "source": [
    "import pprint\n",
    "\n",
    "pprint.pprint(address_latlng)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 5 passos para deduplicação\n",
    "0. Pré-processamento\n",
    "    * input: registros\n",
    "    * output: registros limpos\n",
    "1. Indexação\n",
    "    * output: pares de registros\n",
    "2. Comparação\n",
    "    * output: para cada par, valores de similaridades (vetores)\n",
    "3. Classificação\n",
    "    * output: pares separados em duplicatas e não-duplicatas\n",
    "4. Clusterização\n",
    "    * output: clusters de registros"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 0/4 - Pré-processamento"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Dataset fabricado baseado no dataset [Restaurant](https://www.cs.utexas.edu/users/ml/riddle/data.html):\n",
    "- dataset frequentemente usado por pesquisadores\n",
    "- consiste em 881 registros de restaurantes dos guias Fodor's e Zagat\n",
    "- contem 150 registros duplicados"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "source": [
    "\\* Verifique as mudanças que introduzimos fazendo um `diff restaurant.original.csv restaurant.csv`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Vamos carregar nosso dataset em um **dataframe** Pandas:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "      <th>phone</th>\n",
       "      <th>type</th>\n",
       "      <th>cluster</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>arnie morton's of chicago</td>\n",
       "      <td>435 s. la cienega blv.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>310/246-1501</td>\n",
       "      <td>american</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>arnie morton's of chicago</td>\n",
       "      <td>435 s. la cienega blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>310-246-1501</td>\n",
       "      <td>steakhouses</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>arnie morton</td>\n",
       "      <td>435 s. la cienega boulevard</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>310-246-1501</td>\n",
       "      <td>steakhouses</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>art's delicatessen</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "      <td>818/762-1221</td>\n",
       "      <td>american</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>art's deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "      <td>818-762-1221</td>\n",
       "      <td>delis</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>art's deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>818-762-1221</td>\n",
       "      <td>delis</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>hotel bel-air</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "      <td>310/472-1211</td>\n",
       "      <td>californian</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>bel-air hotel</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "      <td>310-472-1211</td>\n",
       "      <td>californian</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>bel-air</td>\n",
       "      <td>701 stone canyon road</td>\n",
       "      <td>bel air</td>\n",
       "      <td>(310) 472-1211</td>\n",
       "      <td>american</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                        name                         addr         city  \\\n",
       "0  arnie morton's of chicago       435 s. la cienega blv.  los angeles   \n",
       "1  arnie morton's of chicago      435 s. la cienega blvd.  los angeles   \n",
       "2               arnie morton  435 s. la cienega boulevard  los angeles   \n",
       "3         art's delicatessen          12224 ventura blvd.  studio city   \n",
       "4                 art's deli          12224 ventura blvd.  studio city   \n",
       "5                 art's deli          12224 ventura blvd.  los angeles   \n",
       "6              hotel bel-air         701 stone canyon rd.      bel air   \n",
       "7              bel-air hotel         701 stone canyon rd.      bel air   \n",
       "8                    bel-air        701 stone canyon road      bel air   \n",
       "\n",
       "            phone         type  cluster  \n",
       "0    310/246-1501     american        0  \n",
       "1    310-246-1501  steakhouses        0  \n",
       "2    310-246-1501  steakhouses        0  \n",
       "3    818/762-1221     american        1  \n",
       "4    818-762-1221        delis        1  \n",
       "5    818-762-1221        delis        1  \n",
       "6    310/472-1211  californian        2  \n",
       "7    310-472-1211  californian        2  \n",
       "8  (310) 472-1211     american        2  "
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_with_truth = pd.read_csv('restaurant.csv', skip_blank_lines=True)\n",
    "df_with_truth.head(9)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Este dataset vem com os registros deduplicados indicados pelo atributo `cluster`."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para mostrar como fazer a deduplicação em um contexto difícil, vamos remover os atributos `phone` e `type`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>arnie morton's of chicago</td>\n",
       "      <td>435 s. la cienega blv.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>arnie morton's of chicago</td>\n",
       "      <td>435 s. la cienega blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>arnie morton</td>\n",
       "      <td>435 s. la cienega boulevard</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>art's delicatessen</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>art's deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>art's deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>hotel bel-air</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>bel-air hotel</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>bel-air</td>\n",
       "      <td>701 stone canyon road</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                        name                         addr         city\n",
       "0  arnie morton's of chicago       435 s. la cienega blv.  los angeles\n",
       "1  arnie morton's of chicago      435 s. la cienega blvd.  los angeles\n",
       "2               arnie morton  435 s. la cienega boulevard  los angeles\n",
       "3         art's delicatessen          12224 ventura blvd.  studio city\n",
       "4                 art's deli          12224 ventura blvd.  studio city\n",
       "5                 art's deli          12224 ventura blvd.  los angeles\n",
       "6              hotel bel-air         701 stone canyon rd.      bel air\n",
       "7              bel-air hotel         701 stone canyon rd.      bel air\n",
       "8                    bel-air        701 stone canyon road      bel air"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = df_with_truth.drop(columns=['cluster', 'phone', 'type'])\n",
    "df.head(9)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Agora vamos pré-processar os registros. Primeiro, limpando `name`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>435 s. la cienega blv.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>435 s. la cienega blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>arnie morton</td>\n",
       "      <td>435 s. la cienega boulevard</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>art s delicatessen</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>art s deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>art s deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>hotel bel air</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>bel air hotel</td>\n",
       "      <td>701 stone canyon rd.</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>bel air</td>\n",
       "      <td>701 stone canyon road</td>\n",
       "      <td>bel air</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                        name                         addr         city\n",
       "0  arnie morton s of chicago       435 s. la cienega blv.  los angeles\n",
       "1  arnie morton s of chicago      435 s. la cienega blvd.  los angeles\n",
       "2               arnie morton  435 s. la cienega boulevard  los angeles\n",
       "3         art s delicatessen          12224 ventura blvd.  studio city\n",
       "4                 art s deli          12224 ventura blvd.  studio city\n",
       "5                 art s deli          12224 ventura blvd.  los angeles\n",
       "6              hotel bel air         701 stone canyon rd.      bel air\n",
       "7              bel air hotel         701 stone canyon rd.      bel air\n",
       "8                    bel air        701 stone canyon road      bel air"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import re\n",
    "\n",
    "irrelevant_regex = re.compile(r'[^a-z0-9\\s]')\n",
    "multispace_regex = re.compile(r'\\s\\s+')\n",
    "\n",
    "def assign_no_symbols_name(df):\n",
    "    return df.assign(\n",
    "        name=df['name'].str.replace(irrelevant_regex, ' ').str.replace(multispace_regex, ' '))\n",
    "\n",
    "df = assign_no_symbols_name(df)\n",
    "df.head(9)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "E depois geocodificando `addr`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "881 819\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "\n",
    "all_addresses = df['addr'].str.cat(df['city'], sep=', ').values\n",
    "unique_addresses = np.unique(all_addresses)\n",
    "print(len(all_addresses), len(unique_addresses))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Geocoding 1/1\n",
      "Failed to get postal from 11\n",
      "Failed to get latlng from 1\n"
     ]
    }
   ],
   "source": [
    "import os.path\n",
    "import json\n",
    "\n",
    "geocoding_filename = 'address_to_geocoding.json'\n",
    "\n",
    "def geocode_addresses(address_to_geocoding):\n",
    "    remaining_addresses = (\n",
    "        set(unique_addresses) -\n",
    "        set(k for k, v in address_to_geocoding.items() if v is not None and 'lat' in v))\n",
    "\n",
    "    with requests.Session() as session:\n",
    "        for i, address in enumerate(remaining_addresses):\n",
    "            print(f\"Geocoding {i + 1}/{len(remaining_addresses)}\")\n",
    "            geocode_result = geocoder.google(address, session=session)\n",
    "            address_to_geocoding[address] = geocode_result.json\n",
    "\n",
    "        with open(geocoding_filename, 'w') as f:\n",
    "            json.dump(address_to_geocoding, f, indent=4)\n",
    "\n",
    "if not os.path.exists(geocoding_filename):\n",
    "    address_to_geocoding = {}\n",
    "    geocode_addresses(address_to_geocoding)\n",
    "else:\n",
    "    with open(geocoding_filename) as f:\n",
    "        address_to_geocoding = json.load(f)\n",
    "    geocode_addresses(address_to_geocoding)\n",
    " \n",
    "address_to_postal = {\n",
    "    k: v['postal']\n",
    "    for k, v in address_to_geocoding.items()\n",
    "    if v is not None and 'postal' in v\n",
    "}\n",
    "address_to_latlng = {\n",
    "    k: (v['lat'], v['lng'])\n",
    "    for k, v in address_to_geocoding.items()\n",
    "    if v is not None\n",
    "}\n",
    "print(f\"Failed to get postal from {len(address_to_geocoding) - len(address_to_postal)}\")\n",
    "print(f\"Failed to get latlng from {len(address_to_geocoding) - len(address_to_latlng)}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "scrolled": false,
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>435 s. la cienega blv.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.070708</td>\n",
       "      <td>-118.376563</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>435 s. la cienega blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.070708</td>\n",
       "      <td>-118.376563</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>arnie morton</td>\n",
       "      <td>435 s. la cienega boulevard</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.070708</td>\n",
       "      <td>-118.376563</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>art s delicatessen</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "      <td>91604</td>\n",
       "      <td>34.142963</td>\n",
       "      <td>-118.399465</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>art s deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>studio city</td>\n",
       "      <td>91604</td>\n",
       "      <td>34.142963</td>\n",
       "      <td>-118.399465</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>art s deli</td>\n",
       "      <td>12224 ventura blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>91604</td>\n",
       "      <td>34.142963</td>\n",
       "      <td>-118.399465</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                        name                         addr         city postal  \\\n",
       "0  arnie morton s of chicago       435 s. la cienega blv.  los angeles  90048   \n",
       "1  arnie morton s of chicago      435 s. la cienega blvd.  los angeles  90048   \n",
       "2               arnie morton  435 s. la cienega boulevard  los angeles  90048   \n",
       "3         art s delicatessen          12224 ventura blvd.  studio city  91604   \n",
       "4                 art s deli          12224 ventura blvd.  studio city  91604   \n",
       "5                 art s deli          12224 ventura blvd.  los angeles  91604   \n",
       "\n",
       "         lat         lng  \n",
       "0  34.070708 -118.376563  \n",
       "1  34.070708 -118.376563  \n",
       "2  34.070708 -118.376563  \n",
       "3  34.142963 -118.399465  \n",
       "4  34.142963 -118.399465  \n",
       "5  34.142963 -118.399465  "
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def assign_postal_lat_lng(df):\n",
    "    addresses = df['addr'].str.cat(df['city'], sep=', ')\n",
    "    addresses_to_postal = [address_to_postal.get(a) for a in addresses]\n",
    "    addresses_to_lat = [address_to_latlng[a][0] if a in address_to_latlng else None for a in addresses]\n",
    "    addresses_to_lng = [address_to_latlng[a][1] if a in address_to_latlng else None for a in addresses]\n",
    "\n",
    "    return df.assign(postal=addresses_to_postal, lat=addresses_to_lat, lng=addresses_to_lng)\n",
    "\n",
    "df = assign_postal_lat_lng(df)\n",
    "df.head(6)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Assim **limpamos** e **enriquecemos** nossos registros."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 1/4 - Indexação"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Nós temos os registros limpos, agora nós precisamos dos **pares** dos registros para compará-los e encontrar duplicatas.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para os próximos passos, usaremos a biblioteca **recordlinkage**, aka **[Python Record Linkage Toolkit](http://recordlinkage.readthedocs.io/en/stable/)**.\n",
    "\n",
    "recordlinkage suporta **dataframes Pandas**, então não precisaremos adaptar os dados que já temos.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Pra produzir os pares, podemos fazer uma indexação complexa usando a classe `Full`. Como isso, teremos que comparar todos os registros entre si:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:recordlinkage:indexing - performance warning - A full index can result in large number of record pairs.\n",
      "Full index: 881 records, 387640 pairs\n"
     ]
    }
   ],
   "source": [
    "import recordlinkage as rl\n",
    "from recordlinkage.index import Full\n",
    "\n",
    "full_indexer = Full()\n",
    "pairs = full_indexer.index(df)\n",
    "\n",
    "print(f\"Full index: {len(df)} records, {len(pairs)} pairs\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<div><img style=\"display: block; margin: 0 auto;\" src=\"full-indexing.png?\" width=1200></div>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "O problema é que o número de pares cresce muito rápido, i.e., **quadraticamente** na medida que os registros crescem: `len(df) * (len(df) - 1) / 2`\n",
    "\n",
    "Para evitar tantas comparações e perda de tempo, precisamos **indexar**. Precisamos produzir **apenas** os pares que têm uma chance mínima representar duplicatas."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "A forma mais básica de indexar se chama **blocking**. É uma regra de indexação que produz pares **que possuem um valor em comum**."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<div><img style=\"display: block; margin: 0 auto;\" src=\"standard-blocking.png?\" width=1000></div>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "No nosso exemplo, vamos produzir pares com o mesmo valor para **postal** (código postal) usando a classe `Block`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Postal index: 6462 pairs\n"
     ]
    }
   ],
   "source": [
    "from recordlinkage.index import Block\n",
    "\n",
    "postal_indexer = Block('postal')\n",
    "pairs = postal_indexer.index(df)\n",
    "\n",
    "print(f\"Postal index: {len(pairs)} pairs\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "# We'll swap the levels of pairs MultiIndex returned by recordlinkage\n",
    "# to make it simpler to interpret. Here we get the upper triangle of the pairs matrix.\n",
    "pairs = pairs.swaplevel().sortlevel()[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[  0,   1],\n",
       "       [  0,   2],\n",
       "       [  0,  32],\n",
       "       [  0,  33],\n",
       "       [  0,  35],\n",
       "       [  0,  36],\n",
       "       [  0, 262],\n",
       "       [  0, 266],\n",
       "       [  0, 273],\n",
       "       [  0, 293]])"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pairs.to_frame()[:10].values"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>postal</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>90048</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>arnie morton s of chicago</td>\n",
       "      <td>90048</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>arnie morton</td>\n",
       "      <td>90048</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                        name postal\n",
       "0  arnie morton s of chicago  90048\n",
       "1  arnie morton s of chicago  90048\n",
       "2               arnie morton  90048"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df[['name', 'postal']].head(3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Existem muitas outras formas de indexação além de **Blocking**:\n",
    "* Sorted Neighborhood\n",
    "* Canopy Clustering\n",
    "* Locality Sensitive Hashing\n",
    "* ou até... **aprender** como indexar com Machine Learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 2/4 - Comparação"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Com os pares, vamos produzir um **vetor** (feature vector) para cada par.  \n",
    "Cada vetor contem as similaridades dos atributos. Algo como:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>latlng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>100</th>\n",
       "      <th>200</th>\n",
       "      <td>0.5</td>\n",
       "      <td>0.8</td>\n",
       "      <td>0.9</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "         name  addr  postal  latlng\n",
       "100 200   0.5   0.8     0.9       1"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pd.DataFrame([[0.5, 0.8, 0.9, 1]],\n",
    "             columns=['name', 'addr', 'postal', 'latlng'],\n",
    "             index=pd.MultiIndex.from_arrays([[100], [200]]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>latlng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>100</th>\n",
       "      <th>200</th>\n",
       "      <td>0.5</td>\n",
       "      <td>0.8</td>\n",
       "      <td>0.9</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "         name  addr  postal  latlng\n",
       "100 200   0.5   0.8     0.9       1"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pd.DataFrame([[0.5, 0.8, 0.9, 1]],\n",
    "             columns=['name', 'addr', 'postal', 'latlng'],\n",
    "             index=pd.MultiIndex.from_arrays([[100], [200]]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Nesse exemplo, o par `(100, 200)` possui:\n",
    "- Valores pouco parecidos em `name`\n",
    "- Valores parecidos em `addr`\n",
    "- Valores muito parecidos em `postal`\n",
    "- Valores iguais de `latlng`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para computar esse vetor, preciamos definir pelo menos **uma função de similaridade** para cada atributo:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [],
   "source": [
    "comp = rl.Compare()\n",
    "comp.string('name', 'name', method='jarowinkler', label='name')\n",
    "comp.string('addr', 'addr', method='jarowinkler', label='addr')\n",
    "comp.string('postal', 'postal', method='jarowinkler', label='postal')\n",
    "comp.geo('lat', 'lng', 'lat', 'lng', method='exp', scale=0.1, offset=0.01, label='latlng');"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Depois, é só chamar o `compute` com os pares indexados `pairs` e os registros `df` para obter os **vetores**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>latlng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th rowspan=\"5\" valign=\"top\">0</th>\n",
       "      <th>1</th>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.985507</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.896000</td>\n",
       "      <td>0.910774</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>32</th>\n",
       "      <td>0.520128</td>\n",
       "      <td>0.580927</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.000025</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>33</th>\n",
       "      <td>0.556630</td>\n",
       "      <td>0.580927</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.000025</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>35</th>\n",
       "      <td>0.505476</td>\n",
       "      <td>0.528075</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.030131</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "          name      addr  postal    latlng\n",
       "0 1   1.000000  0.985507     1.0  1.000000\n",
       "  2   0.896000  0.910774     1.0  1.000000\n",
       "  32  0.520128  0.580927     1.0  0.000025\n",
       "  33  0.556630  0.580927     1.0  0.000025\n",
       "  35  0.505476  0.528075     1.0  0.030131"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "comparison_vectors = comp.compute(pairs, df)\n",
    "comparison_vectors.head(5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 3/4 - Classificação"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Com nossos vetores, vamos explorar **diferentes formas** de classificá-los como duplicatas ou não."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Classificação por limiar"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Uma maneira simples de classificar os vetores como duplicatas ou não é computar uma **média ponderada** para servir de **pontuação** (score) e usar um **limiar** (threshold):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>latlng</th>\n",
       "      <th>score</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th rowspan=\"5\" valign=\"top\">0</th>\n",
       "      <th>1</th>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.985507</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.996047</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.896000</td>\n",
       "      <td>0.910774</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.928393</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>32</th>\n",
       "      <td>0.520128</td>\n",
       "      <td>0.580927</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.000025</td>\n",
       "      <td>0.485770</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>33</th>\n",
       "      <td>0.556630</td>\n",
       "      <td>0.580927</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.000025</td>\n",
       "      <td>0.502362</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>35</th>\n",
       "      <td>0.505476</td>\n",
       "      <td>0.528075</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.030131</td>\n",
       "      <td>0.470170</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "          name      addr  postal    latlng     score\n",
       "0 1   1.000000  0.985507     1.0  1.000000  0.996047\n",
       "  2   0.896000  0.910774     1.0  1.000000  0.928393\n",
       "  32  0.520128  0.580927     1.0  0.000025  0.485770\n",
       "  33  0.556630  0.580927     1.0  0.000025  0.502362\n",
       "  35  0.505476  0.528075     1.0  0.030131  0.470170"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "scores = np.average(\n",
    "    comparison_vectors.values,\n",
    "    axis=1,\n",
    "    weights=[50, 30, 10, 20])\n",
    "scored_comparison_vectors = comparison_vectors.assign(score=scores)\n",
    "scored_comparison_vectors.head(5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Sabemos que o registro `0` realmente é uma duplicata do `1` e `2`. Para simplificar, vamos **classificar como duplicatas** quaisquer pares com `score >= 0.9`. Este é o nosso **limiar**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>latlng</th>\n",
       "      <th>score</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">0</th>\n",
       "      <th>1</th>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.985507</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.996047</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.896000</td>\n",
       "      <td>0.910774</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.928393</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <td>0.896000</td>\n",
       "      <td>0.923779</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.931940</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th rowspan=\"2\" valign=\"top\">3</th>\n",
       "      <th>4</th>\n",
       "      <td>0.911111</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.959596</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>0.911111</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.959596</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.000000</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "         name      addr  postal  latlng     score\n",
       "0 1  1.000000  0.985507     1.0     1.0  0.996047\n",
       "  2  0.896000  0.910774     1.0     1.0  0.928393\n",
       "1 2  0.896000  0.923779     1.0     1.0  0.931940\n",
       "3 4  0.911111  1.000000     1.0     1.0  0.959596\n",
       "  5  0.911111  1.000000     1.0     1.0  0.959596\n",
       "4 5  1.000000  1.000000     1.0     1.0  1.000000"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "matches = scored_comparison_vectors[\n",
    "    scored_comparison_vectors['score'] >= 0.9]\n",
    "matches.head(6)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Já que temos os registros duplicados marcados pelo atributo `cluster`, podemos **avaliar** o quão bom é o nosso método de classificação por limiar:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Golden pairs: 150\n"
     ]
    }
   ],
   "source": [
    "golden_pairs = Block('cluster').index(df_with_truth)\n",
    "golden_pairs = golden_pairs.swaplevel().sortlevel()[0]\n",
    "print(\"Golden pairs:\", len(golden_pairs))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "true_positives total: 128\n",
      "false_positives total: 2\n",
      "false_negatives total: 22\n"
     ]
    }
   ],
   "source": [
    "found_pairs_set = set(matches.index)\n",
    "\n",
    "golden_pairs_set = set(golden_pairs)\n",
    "\n",
    "true_positives = golden_pairs_set & found_pairs_set\n",
    "false_positives = found_pairs_set - golden_pairs_set\n",
    "false_negatives = golden_pairs_set - found_pairs_set\n",
    "\n",
    "print('true_positives total:', len(true_positives))\n",
    "print('false_positives total:', len(false_positives))\n",
    "print('false_negatives total:', len(false_negatives))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Tivemos um pequeno número de **falso positivos**. São casos complicados, pois são pares de registros praticamente idênticos mas que não representam o mesmo restaurante:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "False positives:\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>196</th>\n",
       "      <td>ritz carlton cafe buckhead</td>\n",
       "      <td>3434 peachtree rd. ne</td>\n",
       "      <td>30326</td>\n",
       "      <td>33.850488</td>\n",
       "      <td>-84.363631</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>198</th>\n",
       "      <td>ritz carlton dining room buckhead</td>\n",
       "      <td>3434 peachtree rd. ne</td>\n",
       "      <td>30326</td>\n",
       "      <td>33.850488</td>\n",
       "      <td>-84.363631</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                   name                   addr postal  \\\n",
       "196         ritz carlton cafe buckhead   3434 peachtree rd. ne  30326   \n",
       "198  ritz carlton dining room buckhead   3434 peachtree rd. ne  30326   \n",
       "\n",
       "           lat        lng  \n",
       "196  33.850488 -84.363631  \n",
       "198  33.850488 -84.363631  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>200</th>\n",
       "      <td>ritz carlton restaurant</td>\n",
       "      <td>181 peachtree st.</td>\n",
       "      <td>30303</td>\n",
       "      <td>33.758572</td>\n",
       "      <td>-84.387181</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>839</th>\n",
       "      <td>ritz carlton cafe atlanta</td>\n",
       "      <td>181 peachtree st.</td>\n",
       "      <td>30303</td>\n",
       "      <td>33.758572</td>\n",
       "      <td>-84.387181</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                           name               addr postal        lat  \\\n",
       "200     ritz carlton restaurant  181 peachtree st.  30303  33.758572   \n",
       "839  ritz carlton cafe atlanta   181 peachtree st.  30303  33.758572   \n",
       "\n",
       "           lng  \n",
       "200 -84.387181  \n",
       "839 -84.387181  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print(f\"False positives:\")\n",
    "for false_positive_pair in false_positives:\n",
    "    display(df.loc[list(false_positive_pair)][['name', 'addr', 'postal', 'lat', 'lng']])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Por outro lado, tivemos um alto número de **falso negativos**. Nosso método por limiar **perdeu** muitas duplicatas!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "False negatives (sample 10 of 22):\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>182</th>\n",
       "      <td>heera of india</td>\n",
       "      <td>595 piedmont ave.  rio shopping mall</td>\n",
       "      <td>30324</td>\n",
       "      <td>33.798336</td>\n",
       "      <td>-84.371044</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>183</th>\n",
       "      <td>heera of india</td>\n",
       "      <td>595 piedmont ave.</td>\n",
       "      <td>30308</td>\n",
       "      <td>33.770298</td>\n",
       "      <td>-84.381122</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "               name                                  addr postal        lat  \\\n",
       "182  heera of india  595 piedmont ave.  rio shopping mall  30324  33.798336   \n",
       "183  heera of india                     595 piedmont ave.  30308  33.770298   \n",
       "\n",
       "           lng  \n",
       "182 -84.371044  \n",
       "183 -84.381122  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>164</th>\n",
       "      <td>abruzzi</td>\n",
       "      <td>2355 peachtree rd.  peachtree battle shopping ...</td>\n",
       "      <td>30305</td>\n",
       "      <td>33.819818</td>\n",
       "      <td>-84.386643</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>165</th>\n",
       "      <td>abruzzi</td>\n",
       "      <td>2355 peachtree rd. ne</td>\n",
       "      <td>30305</td>\n",
       "      <td>33.819166</td>\n",
       "      <td>-84.387518</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        name                                               addr postal  \\\n",
       "164  abruzzi  2355 peachtree rd.  peachtree battle shopping ...  30305   \n",
       "165  abruzzi                              2355 peachtree rd. ne  30305   \n",
       "\n",
       "           lat        lng  \n",
       "164  33.819818 -84.386643  \n",
       "165  33.819166 -84.387518  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>28</th>\n",
       "      <td>restaurant katsu</td>\n",
       "      <td>1972 n. hillhurst ave.</td>\n",
       "      <td>90027</td>\n",
       "      <td>34.107403</td>\n",
       "      <td>-118.287172</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>29</th>\n",
       "      <td>katsu</td>\n",
       "      <td>1972 hillhurst ave.</td>\n",
       "      <td>90027</td>\n",
       "      <td>34.107403</td>\n",
       "      <td>-118.287172</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                name                    addr postal        lat         lng\n",
       "28  restaurant katsu  1972 n. hillhurst ave.  90027  34.107403 -118.287172\n",
       "29             katsu     1972 hillhurst ave.  90027  34.107403 -118.287172"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>36</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w 3rd</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.073421</td>\n",
       "      <td>-118.381098</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>37</th>\n",
       "      <td>locanda</td>\n",
       "      <td>w. third st.</td>\n",
       "      <td>None</td>\n",
       "      <td>34.068947</td>\n",
       "      <td>-118.322599</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "              name          addr postal        lat         lng\n",
       "36  locanda veneta    8638 w 3rd  90048  34.073421 -118.381098\n",
       "37         locanda  w. third st.   None  34.068947 -118.322599"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>40</th>\n",
       "      <td>the palm</td>\n",
       "      <td>9001 santa monica blvd.</td>\n",
       "      <td>90069</td>\n",
       "      <td>34.083473</td>\n",
       "      <td>-118.387373</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>41</th>\n",
       "      <td>palm the los angeles</td>\n",
       "      <td>9001 santa monica blvd.</td>\n",
       "      <td>90069</td>\n",
       "      <td>34.083064</td>\n",
       "      <td>-118.387282</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                     name                     addr postal        lat  \\\n",
       "40               the palm  9001 santa monica blvd.  90069  34.083473   \n",
       "41  palm the los angeles   9001 santa monica blvd.  90069  34.083064   \n",
       "\n",
       "           lng  \n",
       "40 -118.387373  \n",
       "41 -118.387282  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>136</th>\n",
       "      <td>shun lee west</td>\n",
       "      <td>43 w. 65th st.</td>\n",
       "      <td>10023</td>\n",
       "      <td>40.772900</td>\n",
       "      <td>-73.981348</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>137</th>\n",
       "      <td>shun lee palace</td>\n",
       "      <td>155 e. 55th st.</td>\n",
       "      <td>10022</td>\n",
       "      <td>40.759435</td>\n",
       "      <td>-73.969072</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                name             addr postal        lat        lng\n",
       "136    shun lee west   43 w. 65th st.  10023  40.772900 -73.981348\n",
       "137  shun lee palace  155 e. 55th st.  10022  40.759435 -73.969072"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>34</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>3rd st.</td>\n",
       "      <td>None</td>\n",
       "      <td>34.009031</td>\n",
       "      <td>-118.488174</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>35</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w. third st.</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.073421</td>\n",
       "      <td>-118.381098</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "              name               addr postal        lat         lng\n",
       "34  locanda veneta            3rd st.   None  34.009031 -118.488174\n",
       "35  locanda veneta  8638 w. third st.  90048  34.073421 -118.381098"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>110</th>\n",
       "      <td>mesa grill</td>\n",
       "      <td>102 5th ave.  between 15th and 16th sts.</td>\n",
       "      <td>10001</td>\n",
       "      <td>40.748441</td>\n",
       "      <td>-73.985664</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>111</th>\n",
       "      <td>mesa grill</td>\n",
       "      <td>102 fifth ave.</td>\n",
       "      <td>10011</td>\n",
       "      <td>40.737045</td>\n",
       "      <td>-73.993119</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "           name                                      addr postal        lat  \\\n",
       "110  mesa grill  102 5th ave.  between 15th and 16th sts.  10001  40.748441   \n",
       "111  mesa grill                            102 fifth ave.  10011  40.737045   \n",
       "\n",
       "           lng  \n",
       "110 -73.985664  \n",
       "111 -73.993119  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>35</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w. third st.</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.073421</td>\n",
       "      <td>-118.381098</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>37</th>\n",
       "      <td>locanda</td>\n",
       "      <td>w. third st.</td>\n",
       "      <td>None</td>\n",
       "      <td>34.068947</td>\n",
       "      <td>-118.322599</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "              name               addr postal        lat         lng\n",
       "35  locanda veneta  8638 w. third st.  90048  34.073421 -118.381098\n",
       "37         locanda       w. third st.   None  34.068947 -118.322599"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>144</th>\n",
       "      <td>uncle nick s</td>\n",
       "      <td>747 9th ave.  between 50th and 51st sts.</td>\n",
       "      <td>10019</td>\n",
       "      <td>40.763835</td>\n",
       "      <td>-73.988912</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>145</th>\n",
       "      <td>uncle nick s</td>\n",
       "      <td>747 ninth ave.</td>\n",
       "      <td>10019</td>\n",
       "      <td>40.763884</td>\n",
       "      <td>-73.989002</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "             name                                      addr postal        lat  \\\n",
       "144  uncle nick s  747 9th ave.  between 50th and 51st sts.  10019  40.763835   \n",
       "145  uncle nick s                            747 ninth ave.  10019  40.763884   \n",
       "\n",
       "           lng  \n",
       "144 -73.988912  \n",
       "145 -73.989002  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print(f\"False negatives (sample 10 of {len(false_negatives)}):\")\n",
    "for false_negative_pair in list(false_negatives)[:10]:\n",
    "    display(df.loc[list(false_negative_pair)][['name', 'addr', 'postal', 'lat', 'lng']])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Ao invés de tentar **advinhar** um limiar, podemos fazer algo mais inteligente?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Classificação supervisionada"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Podemos treinar um **classificador** capaz de classificar os pares em duplicatas ou não, baseado em um dataset de treinamento que podemos providenciar:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "      <th>cluster</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>3rd st.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w. third st.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w 3rd</td>\n",
       "      <td>st los angeles</td>\n",
       "      <td>13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>cafe lalo</td>\n",
       "      <td>201 w. 83rd st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>26</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>cafe lalo</td>\n",
       "      <td>201 w. 83rd st.</td>\n",
       "      <td>new york city</td>\n",
       "      <td>26</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>les celebrites</td>\n",
       "      <td>160 central park s</td>\n",
       "      <td>new york</td>\n",
       "      <td>42</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>les celebrites</td>\n",
       "      <td>155 w. 58th st.</td>\n",
       "      <td>new york city</td>\n",
       "      <td>42</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>second avenue deli</td>\n",
       "      <td>156 2nd ave. at 10th st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>second avenue deli</td>\n",
       "      <td>156 second ave.</td>\n",
       "      <td>new york city</td>\n",
       "      <td>58</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>smith &amp; wollensky</td>\n",
       "      <td>201 e. 49th st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>62</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>10</th>\n",
       "      <td>smith &amp; wollensky</td>\n",
       "      <td>797 third ave.</td>\n",
       "      <td>new york city</td>\n",
       "      <td>62</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>11</th>\n",
       "      <td>chin's</td>\n",
       "      <td>3200 las vegas blvd. s</td>\n",
       "      <td>las vegas</td>\n",
       "      <td>67</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>12</th>\n",
       "      <td>chin's</td>\n",
       "      <td>3200 las vegas blvd. s.</td>\n",
       "      <td>las vegas</td>\n",
       "      <td>67</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13</th>\n",
       "      <td>toulouse</td>\n",
       "      <td>b peachtree rd.</td>\n",
       "      <td>atlanta</td>\n",
       "      <td>92</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>14</th>\n",
       "      <td>toulouse</td>\n",
       "      <td>293-b peachtree rd.</td>\n",
       "      <td>atlanta</td>\n",
       "      <td>92</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>15</th>\n",
       "      <td>rose pistola</td>\n",
       "      <td>532 columbus ave.</td>\n",
       "      <td>san francisco</td>\n",
       "      <td>111</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>16</th>\n",
       "      <td>rose pistola</td>\n",
       "      <td>532 columbus ave.</td>\n",
       "      <td>san francisco</td>\n",
       "      <td>111</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>17</th>\n",
       "      <td>bistro garden</td>\n",
       "      <td>176 n. canon dr.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>115</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>18</th>\n",
       "      <td>remi</td>\n",
       "      <td>3rd st. promenade</td>\n",
       "      <td>santa monica</td>\n",
       "      <td>159</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>19</th>\n",
       "      <td>remi</td>\n",
       "      <td>145 w. 53rd st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>334</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>20</th>\n",
       "      <td>west</td>\n",
       "      <td>63rd street steakhouse 44 w. 63rd st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>375</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>21</th>\n",
       "      <td>bistro</td>\n",
       "      <td>3400 las vegas blvd. s</td>\n",
       "      <td>las vegas</td>\n",
       "      <td>429</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>22</th>\n",
       "      <td>mikado</td>\n",
       "      <td>3400 las vegas blvd. s</td>\n",
       "      <td>las vegas</td>\n",
       "      <td>446</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>23</th>\n",
       "      <td>l'osteria del forno</td>\n",
       "      <td>519 columbus ave.</td>\n",
       "      <td>san francisco</td>\n",
       "      <td>490</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>24</th>\n",
       "      <td>stars</td>\n",
       "      <td>150 redwood alley</td>\n",
       "      <td>san francisco</td>\n",
       "      <td>514</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>25</th>\n",
       "      <td>stars cafe</td>\n",
       "      <td>500 van ness ave.</td>\n",
       "      <td>san francisco</td>\n",
       "      <td>515</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>26</th>\n",
       "      <td>belvedere  the</td>\n",
       "      <td>9882 little santa monica blvd.</td>\n",
       "      <td>beverly hills</td>\n",
       "      <td>537</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>27</th>\n",
       "      <td>bernard's</td>\n",
       "      <td>515 s. olive st.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>539</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>28</th>\n",
       "      <td>bistro 45</td>\n",
       "      <td>45 s. mentor ave.</td>\n",
       "      <td>pasadena</td>\n",
       "      <td>540</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>29</th>\n",
       "      <td>cafe '50s</td>\n",
       "      <td>838 lincoln blvd.</td>\n",
       "      <td>venice</td>\n",
       "      <td>545</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>30</th>\n",
       "      <td>cafe blanc</td>\n",
       "      <td>9777 little santa monica blvd.</td>\n",
       "      <td>beverly hills</td>\n",
       "      <td>546</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>31</th>\n",
       "      <td>la cachette</td>\n",
       "      <td>10506 little santa monica blvd.</td>\n",
       "      <td>century city</td>\n",
       "      <td>568</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>32</th>\n",
       "      <td>moongate</td>\n",
       "      <td>3400 las vegas blvd. s.</td>\n",
       "      <td>las vegas</td>\n",
       "      <td>666</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                   name                                   addr  \\\n",
       "0        locanda veneta                                3rd st.   \n",
       "1        locanda veneta                      8638 w. third st.   \n",
       "2        locanda veneta                             8638 w 3rd   \n",
       "3             cafe lalo                        201 w. 83rd st.   \n",
       "4             cafe lalo                        201 w. 83rd st.   \n",
       "5        les celebrites                     160 central park s   \n",
       "6        les celebrites                        155 w. 58th st.   \n",
       "7    second avenue deli               156 2nd ave. at 10th st.   \n",
       "8    second avenue deli                        156 second ave.   \n",
       "9     smith & wollensky                        201 e. 49th st.   \n",
       "10    smith & wollensky                         797 third ave.   \n",
       "11               chin's                 3200 las vegas blvd. s   \n",
       "12               chin's                3200 las vegas blvd. s.   \n",
       "13             toulouse                        b peachtree rd.   \n",
       "14             toulouse                    293-b peachtree rd.   \n",
       "15         rose pistola                      532 columbus ave.   \n",
       "16         rose pistola                      532 columbus ave.   \n",
       "17        bistro garden                       176 n. canon dr.   \n",
       "18                 remi                      3rd st. promenade   \n",
       "19                 remi                        145 w. 53rd st.   \n",
       "20                 west  63rd street steakhouse 44 w. 63rd st.   \n",
       "21               bistro                 3400 las vegas blvd. s   \n",
       "22               mikado                 3400 las vegas blvd. s   \n",
       "23  l'osteria del forno                      519 columbus ave.   \n",
       "24                stars                      150 redwood alley   \n",
       "25           stars cafe                      500 van ness ave.   \n",
       "26       belvedere  the         9882 little santa monica blvd.   \n",
       "27            bernard's                       515 s. olive st.   \n",
       "28            bistro 45                      45 s. mentor ave.   \n",
       "29            cafe '50s                      838 lincoln blvd.   \n",
       "30           cafe blanc         9777 little santa monica blvd.   \n",
       "31          la cachette        10506 little santa monica blvd.   \n",
       "32             moongate                3400 las vegas blvd. s.   \n",
       "\n",
       "              city  cluster  \n",
       "0      los angeles       13  \n",
       "1      los angeles       13  \n",
       "2   st los angeles       13  \n",
       "3         new york       26  \n",
       "4    new york city       26  \n",
       "5         new york       42  \n",
       "6    new york city       42  \n",
       "7         new york       58  \n",
       "8    new york city       58  \n",
       "9         new york       62  \n",
       "10   new york city       62  \n",
       "11       las vegas       67  \n",
       "12       las vegas       67  \n",
       "13         atlanta       92  \n",
       "14         atlanta       92  \n",
       "15   san francisco      111  \n",
       "16   san francisco      111  \n",
       "17     los angeles      115  \n",
       "18    santa monica      159  \n",
       "19        new york      334  \n",
       "20        new york      375  \n",
       "21       las vegas      429  \n",
       "22       las vegas      446  \n",
       "23   san francisco      490  \n",
       "24   san francisco      514  \n",
       "25   san francisco      515  \n",
       "26   beverly hills      537  \n",
       "27     los angeles      539  \n",
       "28        pasadena      540  \n",
       "29          venice      545  \n",
       "30   beverly hills      546  \n",
       "31    century city      568  \n",
       "32       las vegas      666  "
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_training = pd.read_csv('restaurant-training.csv', skip_blank_lines=True)\n",
    "df_training = df_training.drop(columns=['type', 'phone'])\n",
    "df_training"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para o classificador aprender corretamente, devemos **pré-processar** nosso dataset de treinamento também:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "      <th>cluster</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>3rd st.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>13</td>\n",
       "      <td>None</td>\n",
       "      <td>34.009031</td>\n",
       "      <td>-118.488174</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w. third st.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>13</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.073421</td>\n",
       "      <td>-118.381098</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>locanda veneta</td>\n",
       "      <td>8638 w 3rd</td>\n",
       "      <td>st los angeles</td>\n",
       "      <td>13</td>\n",
       "      <td>90048</td>\n",
       "      <td>34.073421</td>\n",
       "      <td>-118.381098</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>cafe lalo</td>\n",
       "      <td>201 w. 83rd st.</td>\n",
       "      <td>new york</td>\n",
       "      <td>26</td>\n",
       "      <td>10024</td>\n",
       "      <td>40.785981</td>\n",
       "      <td>-73.976727</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>cafe lalo</td>\n",
       "      <td>201 w. 83rd st.</td>\n",
       "      <td>new york city</td>\n",
       "      <td>26</td>\n",
       "      <td>10024</td>\n",
       "      <td>40.785981</td>\n",
       "      <td>-73.976727</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "             name               addr            city  cluster postal  \\\n",
       "0  locanda veneta            3rd st.     los angeles       13   None   \n",
       "1  locanda veneta  8638 w. third st.     los angeles       13  90048   \n",
       "2  locanda veneta         8638 w 3rd  st los angeles       13  90048   \n",
       "3       cafe lalo    201 w. 83rd st.        new york       26  10024   \n",
       "4       cafe lalo    201 w. 83rd st.   new york city       26  10024   \n",
       "\n",
       "         lat         lng  \n",
       "0  34.009031 -118.488174  \n",
       "1  34.073421 -118.381098  \n",
       "2  34.073421 -118.381098  \n",
       "3  40.785981  -73.976727  \n",
       "4  40.785981  -73.976727  "
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_training = assign_no_symbols_name(df_training)\n",
    "df_training = assign_postal_lat_lng(df_training)\n",
    "df_training.head(5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "O classificador que usaremos é o **Support Vector Machine**. SVMs são bons para deduplicação pois:\n",
    "- São resilientes a ruídos nos dados (pares incorretos no treinamento, colunas faltando, etc.)\n",
    "- Aceitam atributos (features) com similaridades correlacionadas (como 'addr' e 'latlng')\n",
    "- São robustos a desbalanceamento de classes no dataset de treinamento  \n",
    "(e.g. mais não-duplicatas, menos duplicatas, o que é natural em deduplicação) [3]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:recordlinkage:indexing - performance warning - A full index can result in large number of record pairs.\n"
     ]
    }
   ],
   "source": [
    "all_training_pairs = Full().index(df_training)\n",
    "matches_training_pairs = Block('cluster').index(df_training)\n",
    "\n",
    "training_vectors = comp.compute(all_training_pairs, df_training)\n",
    "\n",
    "svm = rl.SVMClassifier()\n",
    "svm.fit(training_vectors, matches_training_pairs);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "true_positives total: 128\n",
      "false_positives total: 2\n",
      "false_negatives total: 22\n",
      "\n",
      "svm_true_positives total: 133\n",
      "svm_false_positives total: 3\n",
      "svm_false_negatives total: 17\n"
     ]
    }
   ],
   "source": [
    "svm_pairs = svm.predict(comparison_vectors)\n",
    "svm_found_pairs_set = set(svm_pairs)\n",
    "\n",
    "svm_true_positives = golden_pairs_set & svm_found_pairs_set\n",
    "svm_false_positives = svm_found_pairs_set - golden_pairs_set\n",
    "svm_false_negatives = golden_pairs_set - svm_found_pairs_set\n",
    "\n",
    "print('true_positives total:', len(true_positives))\n",
    "print('false_positives total:', len(false_positives))\n",
    "print('false_negatives total:', len(false_negatives))\n",
    "print()\n",
    "print('svm_true_positives total:', len(svm_true_positives))\n",
    "print('svm_false_positives total:', len(svm_false_positives))\n",
    "print('svm_false_negatives total:', len(svm_false_negatives))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Um **pouco melhor**, mas pouca diferença..."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Poderíamos tentar [outros classificadores](https://recordlinkage.readthedocs.io/en/stable/notebooks/classifiers.html) disponíveis na biblioteca recordlinkage, mas a verdade é:\n",
    "- É bem difícil **difícil montar um dataset de treinamento** com exemplos suficientes para os várias casos de similaridade de atributos;\n",
    "- É bem difícil **escolher os atributos para blocking**. Talvez estamos:\n",
    "  * introduzindo falso negativos (perdendo duplicatas reais) devido ao blocking (não possuem o mesmo `postal`)\n",
    "  * introduzindo falso positivos (adicionando duplicatas falsas) devido ao blocking + classificador permissivos (possuem mesmo `postal` e são muito parecidas para o classificador)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para lidar com essas **incertezas** temos...\n",
    "\n",
    "### Classificação Active Learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Métodos Active Learning identificam **pares para treinamento** que maximizam a qualidade dos resultados (precision/recall) [3].\n",
    "\n",
    "É possível otimizar tanto as **atributos e regras de blocking** quanto o **classificador**."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Uma biblioteca Python chamada [**Dedupe**](https://docs.dedupe.io/en/latest/index.html) implementa Active Learning. Vamos vê-la na prática:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "import logging; logging.disable(level=logging.NOTSET)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [],
   "source": [
    "from svm_dedupe import SVMDedupe\n",
    "import dedupe\n",
    "\n",
    "fields = [\n",
    "    {\n",
    "        'field': 'name', 'variable name': 'name', 'type': 'JaroWinkler',\n",
    "    },\n",
    "    {\n",
    "        'field': 'addr', 'variable name': 'addr', 'type': 'JaroWinkler',\n",
    "    },\n",
    "    {\n",
    "        'field': 'postal', 'variable name': 'postal', 'type': 'JaroWinkler'\n",
    "    },\n",
    "    {\n",
    "        'field': 'latlng', 'variable name': 'latlng', 'type': 'ExpLatLong'\n",
    "    },\n",
    "]\n",
    "\n",
    "deduper = SVMDedupe(fields)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "INFO:dedupe.api:((SimplePredicate: (firstIntegerPredicate, addr), SimplePredicate: (sameFiveCharStartPredicate, name)), (SimplePredicate: (latLongGridPredicate, latlng), SimplePredicate: (oneGramFingerprint, name)))\n"
     ]
    }
   ],
   "source": [
    "settings_filename = 'dedupe-simple-settings.pickle'\n",
    "if os.path.exists(settings_filename):\n",
    "    with open(settings_filename, 'rb') as sf:\n",
    "        deduper = dedupe.StaticDedupe(sf, num_cores=4)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "source": [
    "Precisamos **adaptar** os dados para o formato que o Dedupe trabalha, pois ele não trabalha com dataframes:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "data_for_dedupe = df.assign(cluster=df_with_truth['cluster']).to_dict('index')\n",
    "for record in data_for_dedupe.values():\n",
    "    # Change nans to None\n",
    "    for k, v in record.items():\n",
    "        if isinstance(v, float) and np.isnan(v):\n",
    "            record[k] = None\n",
    "    \n",
    "    # Move lat and lng to a single field latlng\n",
    "    lat = record.pop('lat')\n",
    "    lng = record.pop('lng')\n",
    "    if lat is not None and lng is not None:\n",
    "        record['latlng'] = (lat, lng)\n",
    "    else:\n",
    "        record['latlng'] = None"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "O processo de Active Learning consiste em iterativamente **marcar** como duplicata ou não-duplicata cada par que o Dedupe quer aprender. São esses os pares que mais o ajudarão a escolher as regras de blocking e treinar o classificador."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "name : philippe the original\n",
      "addr : 1001 north alameda\n",
      "postal : 90012\n",
      "latlng : (34.059721, -118.237025)\n",
      "\n",
      "name : pisces\n",
      "addr : 95 ave. a\n",
      "postal : 10009\n",
      "latlng : (40.7256332, -73.984031)\n",
      "\n",
      "0/10 positive, 0/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : philippe the original\n",
      "addr : 1001 n. alameda st.\n",
      "postal : 90012\n",
      "latlng : (34.059721, -118.237025)\n",
      "\n",
      "name : mon kee seafood restaurant\n",
      "addr : 679 n. spring st.\n",
      "postal : 90012\n",
      "latlng : (34.0595568, -118.2382488)\n",
      "\n",
      "0/10 positive, 1/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : caffe vivaldi\n",
      "addr : 32 jones st. at bleecker st.\n",
      "postal : 10014\n",
      "latlng : (40.7317316, -74.00298049999999)\n",
      "\n",
      "name : patria\n",
      "addr : 250 park ave. s at 20th st.\n",
      "postal : 10003\n",
      "latlng : (40.7382552, -73.988214)\n",
      "\n",
      "0/10 positive, 2/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : i trulli\n",
      "addr : 122 e. 27th st.  between lexington and park aves.\n",
      "postal : 10028\n",
      "latlng : (40.77961699999999, -73.95631999999999)\n",
      "\n",
      "name : otabe\n",
      "addr : 68 e. 56th st.\n",
      "postal : 10022\n",
      "latlng : (40.7611775, -73.9720541)\n",
      "\n",
      "0/10 positive, 3/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : viva mercado s\n",
      "addr : 6182 w. flamingo rd.\n",
      "postal : 89103\n",
      "latlng : (36.1149027, -115.2269398)\n",
      "\n",
      "name : cafe con leche\n",
      "addr : 424 amsterdam ave.\n",
      "postal : 10024\n",
      "latlng : (40.7841454, -73.9778061)\n",
      "\n",
      "0/10 positive, 4/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : faz\n",
      "addr : 161 sutter st.\n",
      "postal : 94104\n",
      "latlng : (37.78973999999999, -122.4032937)\n",
      "\n",
      "name : splendido embarcadero\n",
      "addr : 4\n",
      "postal : None\n",
      "latlng : (37.7779649, -122.3962019)\n",
      "\n",
      "0/10 positive, 5/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : le bernardin\n",
      "addr : 155 w. 51st st.\n",
      "postal : 10019\n",
      "latlng : (40.7615691, -73.98180479999999)\n",
      "\n",
      "name : le bernardin\n",
      "addr : 155 w. 51st st.\n",
      "postal : 10019\n",
      "latlng : (40.7615691, -73.98180479999999)\n",
      "\n",
      "0/10 positive, 6/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "y\n",
      "\n",
      "name : cafe lalo\n",
      "addr : 201 w. 83rd st.\n",
      "postal : 10024\n",
      "latlng : (40.78598119999999, -73.97672659999999)\n",
      "\n",
      "name : cafe lalo\n",
      "addr : 201 w. 83rd st.\n",
      "postal : 10024\n",
      "latlng : (40.78598119999999, -73.97672659999999)\n",
      "\n",
      "1/10 positive, 6/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "y\n",
      "\n",
      "INFO:dedupe.training:Final predicate set:\n",
      "INFO:dedupe.training:(SimplePredicate: (fingerprint, addr), SimplePredicate: (wholeFieldPredicate, name))\n",
      "name : hi life restaurant and lounge\n",
      "addr : 1340 1st ave. at 72nd st.\n",
      "postal : 10021\n",
      "latlng : (40.7675807, -73.95580029999999)\n",
      "\n",
      "name : trattoria dell arte\n",
      "addr : 900 7th ave.  between 56th and 57th sts.\n",
      "postal : 10106\n",
      "latlng : (40.7654454, -73.9805358)\n",
      "\n",
      "2/10 positive, 6/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : lattanzi ristorante\n",
      "addr : 361 w. 46th st.\n",
      "postal : 10036\n",
      "latlng : (40.7608494, -73.990016)\n",
      "\n",
      "name : pomaire\n",
      "addr : 371 w. 46th st.  off 9th ave.\n",
      "postal : 10036\n",
      "latlng : (40.7609632, -73.9902681)\n",
      "\n",
      "2/10 positive, 7/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : empire korea\n",
      "addr : 6 e. 32nd st.\n",
      "postal : 10016\n",
      "latlng : (40.7465392, -73.9849772)\n",
      "\n",
      "name : tu lan\n",
      "addr : 8 sixth st.\n",
      "postal : 94103\n",
      "latlng : (37.7818759, -122.41013)\n",
      "\n",
      "2/10 positive, 8/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : aquavit\n",
      "addr : 13 w. 54th st.\n",
      "postal : 10019\n",
      "latlng : (40.7616767, -73.976345)\n",
      "\n",
      "name : aquavit\n",
      "addr : 13 w. 54th st.\n",
      "postal : 10019\n",
      "latlng : (40.7616767, -73.976345)\n",
      "\n",
      "2/10 positive, 9/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "y\n",
      "\n",
      "name : mesa grill\n",
      "addr : 102 5th ave.  between 15th and 16th sts.\n",
      "postal : 10011\n",
      "latlng : (40.7370445, -73.9931189)\n",
      "\n",
      "name : fiorello s roman cafe\n",
      "addr : 1900 broadway  between 63rd and 64th sts.\n",
      "postal : 10023\n",
      "latlng : (40.7715867, -73.98138519999999)\n",
      "\n",
      "3/10 positive, 9/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : second avenue deli\n",
      "addr : 156 2nd ave. at 10th st.\n",
      "postal : 10003\n",
      "latlng : (40.7296096, -73.9867012)\n",
      "\n",
      "name : second avenue deli\n",
      "addr : 156 second ave.\n",
      "postal : 10003\n",
      "latlng : (40.7296096, -73.9867012)\n",
      "\n",
      "3/10 positive, 10/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "y\n",
      "\n",
      "name : second street grill\n",
      "addr : 200 e. fremont st.\n",
      "postal : 89101\n",
      "latlng : (36.1709271, -115.1431516)\n",
      "\n",
      "name : paty s\n",
      "addr : 10001 riverside dr.\n",
      "postal : 91602\n",
      "latlng : (34.1524376, -118.3496191)\n",
      "\n",
      "4/10 positive, 10/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "INFO:dedupe.training:Final predicate set:\n",
      "INFO:dedupe.training:(SimplePredicate: (firstIntegerPredicate, addr), SimplePredicate: (wholeFieldPredicate, name))\n",
      "name : cafe ritz carlton buckhead\n",
      "addr : 3434 peachtree rd.\n",
      "postal : 30326\n",
      "latlng : (33.8508073, -84.364227)\n",
      "\n",
      "name : ritz carlton cafe buckhead \n",
      "addr : 3434 peachtree rd. ne\n",
      "postal : 30326\n",
      "latlng : (33.8508073, -84.364227)\n",
      "\n",
      "4/10 positive, 11/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "y\n",
      "\n",
      "INFO:dedupe.training:Final predicate set:\n",
      "INFO:dedupe.training:(SimplePredicate: (fingerprint, name), SimplePredicate: (firstIntegerPredicate, addr))\n",
      "name : le colonial\n",
      "addr : 149 e. 57th st.\n",
      "postal : 10022\n",
      "latlng : (40.7608569, -73.9683494)\n",
      "\n",
      "name : cassell s\n",
      "addr : 3266 w. sixth st.\n",
      "postal : 90020\n",
      "latlng : (34.0634809, -118.2936285)\n",
      "\n",
      "5/10 positive, 11/10 negative\n",
      "Do these records refer to the same thing?\n",
      "(y)es / (n)o / (u)nsure / (f)inished\n",
      "n\n",
      "\n",
      "name : cafe ritz carlton buckhead\n",
      "...\n"
     ]
    }
   ],
   "source": [
    "#     deduper.sample(data_for_dedupe)  \n",
    "#     dedupe.consoleLabel(deduper)\n",
    "#     deduper.train(index_predicates=False)\n",
    "\n",
    "with open('training-simple-input-output.txt') as t:\n",
    "    print('\\n'.join(t.read().split('\\n')[:262]))\n",
    "    print('...')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "source": [
    "Você pode checar o treinamento completo em `training-simple-input-output.txt`.\n",
    "\n",
    "Se você quiser treinar novamente, faça um `rm dedupe-simple-settings.pickle dedupe-simple-training.json`, descomente as linhas acima, e rode tudo de novo desde a inicialização do `deduper`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "from dedupe.convenience import unique\n",
    "\n",
    "def auto_label(deduper):\n",
    "    # based on: dedupe.convenience.consoleLabel\n",
    "    fields = unique(field.field\n",
    "                    for field\n",
    "                    in deduper.data_model.primary_fields)\n",
    "    n_match = 0\n",
    "    n_distinct = 0\n",
    "    buffer_len = 1\n",
    "    examples_buffer = []\n",
    "    uncertain_pairs = []\n",
    "    \n",
    "    while n_match < 12 or n_distinct < 20:\n",
    "        if not uncertain_pairs:\n",
    "            uncertain_pairs = deduper.uncertainPairs()\n",
    "        record_pair = uncertain_pairs.pop()\n",
    "        \n",
    "        n_match = (len(deduper.training_pairs['match']) +\n",
    "                   sum(label == 'match' for _, label in examples_buffer))\n",
    "        n_distinct = (len(deduper.training_pairs['distinct']) +\n",
    "                      sum(label == 'distinct' for _, label in examples_buffer))\n",
    "        \n",
    "        for pair in record_pair:\n",
    "            for field in fields:\n",
    "                line = \"%s : %s\" % (field, pair[field])\n",
    "                print(line)\n",
    "            print()\n",
    "        print(f\"{n_match}/10 positive, {n_distinct}/10 negative\")\n",
    "        print(\"Do these records refer to the same thing?\")\n",
    "        print(\"(y)es / (n)o / (u)nsure / (f)inished\")\n",
    "        \n",
    "        if record_pair[0]['cluster'] == record_pair[1]['cluster']:\n",
    "            print('y')\n",
    "            examples_buffer.insert(0, (record_pair, 'match'))\n",
    "        else:\n",
    "            print('n')\n",
    "            examples_buffer.insert(0, (record_pair, 'distinct'))\n",
    "        print()\n",
    "\n",
    "        if len(examples_buffer) > buffer_len:\n",
    "            record_pair, label = examples_buffer.pop()\n",
    "            if label in ['distinct', 'match']:\n",
    "                examples = {'distinct': [], 'match': []}\n",
    "                examples[label].append(record_pair)\n",
    "                deduper.markPairs(examples)\n",
    "    \n",
    "    for record_pair, label in examples_buffer:\n",
    "        if label in ['distinct', 'match']:\n",
    "            examples = {'distinct': [], 'match': []}\n",
    "            examples[label].append(record_pair)\n",
    "            deduper.markPairs(examples)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [],
   "source": [
    "if not isinstance(deduper, dedupe.StaticDedupe):\n",
    "    deduper.sample(data_for_dedupe)\n",
    "    \n",
    "    training_filename = 'dedupe-simple-training.json'\n",
    "    if os.path.exists(training_filename):\n",
    "        with open(training_filename) as tf:\n",
    "            deduper.readTraining(tf)\n",
    "\n",
    "#     dedupe.consoleLabel(deduper)  # se quiser treinar você mesmo, descomente essa linha e comente a linha abaixo\n",
    "    auto_label(deduper)\n",
    "    \n",
    "    with open(training_filename, 'w') as tf:\n",
    "        deduper.writeTraining(tf)\n",
    "    \n",
    "    deduper.train(index_predicates=False)\n",
    "    \n",
    "    with open(settings_filename, 'wb') as sf:\n",
    "        deduper.writeSettings(sf)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Depois do processo de Active Learning, podemos ver quais **regras de blocking** o `deduper` aprendeu. É bom checar para ver se faz sentido:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((SimplePredicate: (firstIntegerPredicate, addr),\n",
       "  SimplePredicate: (sameFiveCharStartPredicate, name)),\n",
       " (SimplePredicate: (latLongGridPredicate, latlng),\n",
       "  SimplePredicate: (oneGramFingerprint, name)))"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deduper.predicates"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "source": [
    "Essas foram as regras que o `deduper` aprendeu desta longa lista de **possíveis regras**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {
    "scrolled": false,
    "slideshow": {
     "slide_type": "skip"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{SimplePredicate: (alphaNumericPredicate, addr),\n",
       " SimplePredicate: (alphaNumericPredicate, name),\n",
       " SimplePredicate: (alphaNumericPredicate, postal),\n",
       " SimplePredicate: (commonFourGram, addr),\n",
       " SimplePredicate: (commonFourGram, name),\n",
       " SimplePredicate: (commonFourGram, postal),\n",
       " SimplePredicate: (commonIntegerPredicate, addr),\n",
       " SimplePredicate: (commonIntegerPredicate, name),\n",
       " SimplePredicate: (commonIntegerPredicate, postal),\n",
       " SimplePredicate: (commonSixGram, addr),\n",
       " SimplePredicate: (commonSixGram, name),\n",
       " SimplePredicate: (commonSixGram, postal),\n",
       " SimplePredicate: (commonThreeTokens, addr),\n",
       " SimplePredicate: (commonThreeTokens, name),\n",
       " SimplePredicate: (commonThreeTokens, postal),\n",
       " SimplePredicate: (commonTwoTokens, addr),\n",
       " SimplePredicate: (commonTwoTokens, name),\n",
       " SimplePredicate: (commonTwoTokens, postal),\n",
       " SimplePredicate: (doubleMetaphone, addr),\n",
       " SimplePredicate: (doubleMetaphone, name),\n",
       " SimplePredicate: (doubleMetaphone, postal),\n",
       " SimplePredicate: (fingerprint, addr),\n",
       " SimplePredicate: (fingerprint, name),\n",
       " SimplePredicate: (fingerprint, postal),\n",
       " SimplePredicate: (firstIntegerPredicate, addr),\n",
       " SimplePredicate: (firstIntegerPredicate, name),\n",
       " SimplePredicate: (firstIntegerPredicate, postal),\n",
       " SimplePredicate: (firstTokenPredicate, addr),\n",
       " SimplePredicate: (firstTokenPredicate, name),\n",
       " SimplePredicate: (firstTokenPredicate, postal),\n",
       " SimplePredicate: (hundredIntegerPredicate, addr),\n",
       " SimplePredicate: (hundredIntegerPredicate, name),\n",
       " SimplePredicate: (hundredIntegerPredicate, postal),\n",
       " SimplePredicate: (hundredIntegersOddPredicate, addr),\n",
       " SimplePredicate: (hundredIntegersOddPredicate, name),\n",
       " SimplePredicate: (hundredIntegersOddPredicate, postal),\n",
       " SimplePredicate: (latLongGridPredicate, latlng),\n",
       " SimplePredicate: (metaphoneToken, addr),\n",
       " SimplePredicate: (metaphoneToken, name),\n",
       " SimplePredicate: (metaphoneToken, postal),\n",
       " SimplePredicate: (nearIntegersPredicate, addr),\n",
       " SimplePredicate: (nearIntegersPredicate, name),\n",
       " SimplePredicate: (nearIntegersPredicate, postal),\n",
       " SimplePredicate: (oneGramFingerprint, addr),\n",
       " SimplePredicate: (oneGramFingerprint, name),\n",
       " SimplePredicate: (oneGramFingerprint, postal),\n",
       " SimplePredicate: (sameFiveCharStartPredicate, addr),\n",
       " SimplePredicate: (sameFiveCharStartPredicate, name),\n",
       " SimplePredicate: (sameFiveCharStartPredicate, postal),\n",
       " SimplePredicate: (sameSevenCharStartPredicate, addr),\n",
       " SimplePredicate: (sameSevenCharStartPredicate, name),\n",
       " SimplePredicate: (sameSevenCharStartPredicate, postal),\n",
       " SimplePredicate: (sameThreeCharStartPredicate, addr),\n",
       " SimplePredicate: (sameThreeCharStartPredicate, name),\n",
       " SimplePredicate: (sameThreeCharStartPredicate, postal),\n",
       " SimplePredicate: (sortedAcronym, addr),\n",
       " SimplePredicate: (sortedAcronym, name),\n",
       " SimplePredicate: (sortedAcronym, postal),\n",
       " SimplePredicate: (suffixArray, addr),\n",
       " SimplePredicate: (suffixArray, name),\n",
       " SimplePredicate: (suffixArray, postal),\n",
       " SimplePredicate: (tokenFieldPredicate, addr),\n",
       " SimplePredicate: (tokenFieldPredicate, name),\n",
       " SimplePredicate: (tokenFieldPredicate, postal),\n",
       " SimplePredicate: (twoGramFingerprint, addr),\n",
       " SimplePredicate: (twoGramFingerprint, name),\n",
       " SimplePredicate: (twoGramFingerprint, postal),\n",
       " SimplePredicate: (wholeFieldPredicate, addr),\n",
       " SimplePredicate: (wholeFieldPredicate, name),\n",
       " SimplePredicate: (wholeFieldPredicate, postal)}"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deduper.data_model.predicates()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Para terminar o processo de deduplicação com o `deduper`, computamos o **clustering threshold** (limiar) e chamamos **match**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "INFO:dedupe.api:Maximum expected recall and precision\n",
      "INFO:dedupe.api:recall: 0.999\n",
      "INFO:dedupe.api:precision: 0.707\n",
      "INFO:dedupe.api:With threshold: 0.209\n"
     ]
    }
   ],
   "source": [
    "import itertools\n",
    "\n",
    "threshold = deduper.threshold(data_for_dedupe, recall_weight=2)\n",
    "clustered_dupes = deduper.match(data_for_dedupe, threshold=threshold)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Vamos avaliar como o Dedupe performou. Um pequeno **ajuste** é necessário, pois o Dedupe já retorna os clusters ao invés dos pares:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [],
   "source": [
    "dedupe_found_pairs_set = set()\n",
    "for cluster, __ in clustered_dupes:  # we'll explain that later\n",
    "    for pair in itertools.combinations(cluster, 2):\n",
    "        dedupe_found_pairs_set.add(tuple(pair))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {
    "scrolled": true,
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "true_positives total: 128\n",
      "false_positives total: 2\n",
      "false_negatives total: 22\n",
      "\n",
      "svm_true_positives total: 133\n",
      "svm_false_positives total: 3\n",
      "svm_false_negatives total: 17\n",
      "\n",
      "dedupe_true_positives total: 140\n",
      "dedupe_false_positives total: 2\n",
      "dedupe_false_negatives total: 10\n"
     ]
    }
   ],
   "source": [
    "dedupe_true_positives = golden_pairs_set & dedupe_found_pairs_set\n",
    "dedupe_false_positives = dedupe_found_pairs_set - golden_pairs_set\n",
    "dedupe_false_negatives = golden_pairs_set - dedupe_found_pairs_set\n",
    "\n",
    "print('true_positives total:', len(true_positives))\n",
    "print('false_positives total:', len(false_positives))\n",
    "print('false_negatives total:', len(false_negatives))\n",
    "print()\n",
    "print('svm_true_positives total:', len(svm_true_positives))\n",
    "print('svm_false_positives total:', len(svm_false_positives))\n",
    "print('svm_false_negatives total:', len(svm_false_negatives))\n",
    "print()\n",
    "print('dedupe_true_positives total:', len(dedupe_true_positives))\n",
    "print('dedupe_false_positives total:', len(dedupe_false_positives))\n",
    "print('dedupe_false_negatives total:', len(dedupe_false_negatives))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## 4/4 - Clusterização"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "O recordlinkage retorna os **pares de duplicatas**. Mas o Dedupe vai além e retorna os **clusters de duplicatas**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[((0, 1, 2), array([0.71667973, 0.71725396, 0.68507793])),\n",
       " ((3, 4, 5), array([0.70119166, 0.726287  , 0.726287  ])),\n",
       " ((6, 7, 8), array([0.19940663, 0.56712872, 0.26079245])),\n",
       " ((9, 10, 11, 12), array([0.75352792, 0.75352792, 0.75352792, 0.75272876])),\n",
       " ((13, 14), (0.7539285, 0.7539285))]"
      ]
     },
     "execution_count": 48,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "clustered_dupes[:5]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Dedupe finalizou o processo pois **combinou os pares de duplicatas em clusters**. Isso é essencial, pois o seguinte problema pode ocorrer:\n",
    "- Temos os registros `A`, `B`, e `C`\n",
    "- Após deduplicar, temos os pares:\n",
    "- `(A, B)` duplicata\n",
    "- `(B, C)` duplicata\n",
    "- `(A, C)` não-duplicata\n",
    "- E isso **não faz sentido**!\n",
    "\n",
    "A solução a essa ambiguidade se chama computar a **Transitive Closure** através de **clustering**."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Usando alguns métodos privados do Dedupe, é possível computar os **pares de duplicatas** (sem clustering):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [],
   "source": [
    "from dedupe.core import scoreDuplicates\n",
    "\n",
    "candidate_records = itertools.chain.from_iterable(deduper._blockedPairs(deduper._blockData(data_for_dedupe)))\n",
    "dedupe_matches = scoreDuplicates(candidate_records,\n",
    "                                 deduper.data_model,\n",
    "                                 deduper.classifier,\n",
    "                                 deduper.num_cores)\n",
    "dedupe_unclustered_pairs_score_dict = {tuple(pair): score for (pair, score) in dedupe_matches}\n",
    "dedupe_unclustered_found_pairs_set = set(dedupe_unclustered_pairs_score_dict.keys())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Vamos avaliar os pares **unclustered** contra os **clustered**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "dedupe_true_positives total: 140\n",
      "dedupe_false_positives total: 2\n",
      "dedupe_false_negatives total: 10\n",
      "\n",
      "dedupe_unclustered_true_positives total: 138\n",
      "dedupe_unclustered_false_positives total: 5\n",
      "dedupe_unclustered_false_negatives total: 12\n"
     ]
    }
   ],
   "source": [
    "dedupe_unclustered_true_positives = golden_pairs_set & dedupe_unclustered_found_pairs_set\n",
    "dedupe_unclustered_false_positives = dedupe_unclustered_found_pairs_set - golden_pairs_set\n",
    "dedupe_unclustered_false_negatives = golden_pairs_set - dedupe_unclustered_found_pairs_set\n",
    "\n",
    "print('dedupe_true_positives total:', len(dedupe_true_positives))\n",
    "print('dedupe_false_positives total:', len(dedupe_false_positives))\n",
    "print('dedupe_false_negatives total:', len(dedupe_false_negatives))\n",
    "print()\n",
    "print('dedupe_unclustered_true_positives total:', len(dedupe_unclustered_true_positives))\n",
    "print('dedupe_unclustered_false_positives total:', len(dedupe_unclustered_false_positives))\n",
    "print('dedupe_unclustered_false_negatives total:', len(dedupe_unclustered_false_negatives))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Pares Unclustered `!=` Pares Clustered\n",
    "\n",
    "O processo de clustering pode **criar duplicatas** e **remover duplicatas** para resolver ambiguidades."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{(6, 8), (40, 41), (120, 471), (121, 471), (196, 198), (199, 839), (487, 775)}"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "diff_set = dedupe_found_pairs_set ^ dedupe_unclustered_found_pairs_set\n",
    "display(diff_set)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Aqui está um exemplo onde o clustering cria um novo par e **resolve uma ambiguidade**:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {
    "scrolled": false,
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>name</th>\n",
       "      <th>addr</th>\n",
       "      <th>city</th>\n",
       "      <th>postal</th>\n",
       "      <th>lat</th>\n",
       "      <th>lng</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>40</th>\n",
       "      <td>the palm</td>\n",
       "      <td>9001 santa monica blvd.</td>\n",
       "      <td>los angeles</td>\n",
       "      <td>90069</td>\n",
       "      <td>34.083473</td>\n",
       "      <td>-118.387373</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>41</th>\n",
       "      <td>palm the los angeles</td>\n",
       "      <td>9001 santa monica blvd.</td>\n",
       "      <td>w. hollywood</td>\n",
       "      <td>90069</td>\n",
       "      <td>34.083064</td>\n",
       "      <td>-118.387282</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>42</th>\n",
       "      <td>the palm los angeles</td>\n",
       "      <td>9001 sta monica boulevard</td>\n",
       "      <td>hollywood</td>\n",
       "      <td>90069</td>\n",
       "      <td>34.083064</td>\n",
       "      <td>-118.387282</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                     name                       addr          city postal  \\\n",
       "40               the palm    9001 santa monica blvd.   los angeles  90069   \n",
       "41  palm the los angeles     9001 santa monica blvd.  w. hollywood  90069   \n",
       "42   the palm los angeles  9001 sta monica boulevard     hollywood  90069   \n",
       "\n",
       "          lat         lng  \n",
       "40  34.083473 -118.387373  \n",
       "41  34.083064 -118.387282  \n",
       "42  34.083064 -118.387282  "
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x432 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x432 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x432 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from graph_utils import show_cluster_graphs\n",
    "\n",
    "diff_ids = {40, 41, 42}\n",
    "show_cluster_graphs(\n",
    "    df,\n",
    "    golden_pairs_set, dedupe_found_pairs_set, dedupe_unclustered_found_pairs_set,\n",
    "    dedupe_unclustered_pairs_score_dict,\n",
    "    diff_ids)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Próximos Passos"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Procure aprender sobre outros métodos de pré-processamentro, indexação, funções de similaridade, classificadores, etc. Também verifique outros processos como \"Probabilitic Record Linkage\".\n",
    "\n",
    "É interessante começar pela palestra **The Art and Science of Data Matching** [1] e pelo livro **Data Matching** [2]."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Tendo os clusters, como consolidar os dados de vários registros duplicados em um só? Ver **Data Fusion**:\n",
    "  - [Open issue at Python Record Linkage Toolkit](https://github.com/J535D165/recordlinkage/issues/27)\n",
    "  - Christen, 6.12 Merging Matches [2]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "E se novos registros forem adicionados ao dataset? Como lidar com os clusters já formados? Ver **Incremental Record Linkage**:\n",
    "  - A abordagem recomendada pela biblioteca Dedupe é adicionar a um cluster existente ou criar um novo cluster. Para isso, usar a classe [Gazetteer](https://docs.dedupe.io/en/latest/API-documentation.html#gazetteer-objects).\n",
    "  - Outras abordagens, ver os papers:\n",
    "      * [Online Correlation Clustering](https://arxiv.org/pdf/1001.0920.pdf)\n",
    "      * [Incremental Record Linkage](http://www.vldb.org/pvldb/vol7/p697-gruenheid.pdf)\n",
    "      * [Heuristic-based approaches for speeding up incremental record linkage](https://www.sciencedirect.com/science/article/pii/S0164121217302972)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Referências"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "- **Referenciados** nos slides:\n",
    "    * [1] [Mike Mull, The Art and Science of Data Matching](https://www.youtube.com/watch?v=Y-nYEOgq3YE) (slides [here](https://github.com/mikemull/Notebooks/blob/master/PyDataNYCSlides.ipynb))\n",
    "    * [2] [Peter Christen, Data Matching](http://www.springer.com/us/book/9783642311635)\n",
    "    * [3] [Mikhail Yuryevich Bilenko, Learnable Similarity Functions and Their Application to Record Linkage and Clustering](http://www.cs.utexas.edu/~ml/papers/marlin-dissertation-06.pdf)\n",
    "- Outras **talks** sobre Record Linkage com Python:\n",
    "    * [Forest Gregg, Machine Learning and Deduplication](https://www.youtube.com/watch?v=3kGKQqNVibQ)\n",
    "    * [Andrew Rowe, Big Data Deduplication and data matching using Python](https://www.youtube.com/watch?v=Z6mlvrYEYnk)\n",
    "    * [Rhydwyn Mcguire, Record linkage: Join for real life](https://www.youtube.com/watch?v=cEcVIjyHfiQ)\n",
    "- [Basics of Entity Resolution with Python and Dedupe](https://medium.com/district-data-labs/basics-of-entity-resolution-with-python-and-dedupe-bc87440b64d4)\n",
    "- [Automating your Data Cleanup with Python](https://blog.kjamistan.com/automating-your-data-cleanup-with-python/)\n",
    "- [Comparison of Data Matching software](https://github.com/J535D165/data-matching-software)\n",
    "- [Compilation of Record Linkage resources](https://github.com/ropeladder/record-linkage-resources)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "# Obrigado!\n",
    "flavio@vinta.com.br  \n",
    "[@flaviojuvenal](https://twitter.com/flaviojuvenal)  \n",
    "[vinta.com.br](https://www.vinta.com.br?dedupe-slides=1)\n",
    "\n",
    "Agradecimentos especiais a Russell Keith-Magee [@freakboy3742](https://twitter.com/freakboy3742), Forest Timothy Gregg [@forestgregg](https://twitter.com/forestgregg), e Jonathan de Bruin [@J535D165](https://github.com/J535D165)."
   ]
  }
 ],
 "metadata": {
  "celltoolbar": "Slideshow",
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}