From a54a175f9bc9ee08cbd793572e338c075977e3c0 Mon Sep 17 00:00:00 2001 From: ZiyaoGeng <593947521@qq.com> Date: Fri, 8 Apr 2022 16:46:36 +0800 Subject: [PATCH] add docs --- docs/README.md | 4 + docs/data/split_data.md | 236 ++++++++++++++++++++++++++++++++++++++++ docs/exeperiment.md | 18 +++ 3 files changed, 258 insertions(+) create mode 100644 docs/data/split_data.md create mode 100644 docs/exeperiment.md diff --git a/docs/README.md b/docs/README.md index e69de29..900860b 100644 --- a/docs/README.md +++ b/docs/README.md @@ -0,0 +1,4 @@ +# Reclearn文档指南 + +本篇文档是为了方便使用者对Reclearn项目的使用与理解。 + diff --git a/docs/data/split_data.md b/docs/data/split_data.md new file mode 100644 index 0000000..2a31f7d --- /dev/null +++ b/docs/data/split_data.md @@ -0,0 +1,236 @@ +# 数据划分 + +在算法模型的实验中，首先需要完成所需数据集的完成。`reclearn`在`reclearn/data/datasets`中提供了ml-1m、beauty、games、steam、criteo数据集的划分方法。当然读者也可以自己设计方案，本文主要对基本的方法进行简要说明。 + +数据划分主要完成以下任务： + +1. user、item索引的重新编码； +2. 训练集、验证集、测试集的划分，分别进行存储； +3. 记录user、item重新编码后的最大数值，为了指定embedding table的大小； + +ml-1m、beauty、games、steam数据集划分方式相差不大，统一进行说明。而Criteo由于其庞大的数据量，需要单独说明。 + + + +## Moivelens等 + +上述我们已经阐明了主要的任务，但在这基础上还需要注意以下几点： + +- 编码需要从1开始，0用于序列的填充。 +- 取每个用户最后一次的交互行为作为测试集，最后第二次的行为作为验证集，其他作为训练集。 + +因此，例如`data/datasets/movielens.py`中的`split_data`方法： + +```python +def split_data(file_path): + """split movielens for general recommendation + Args: + :param file_path: A string. The file path of 'ratings.dat'. + :return: train_path, val_path, test_path, meta_path + """ + dst_path = os.path.dirname(file_path) + train_path = os.path.join(dst_path, "ml_train.txt") + val_path = os.path.join(dst_path, "ml_val.txt") + test_path = os.path.join(dst_path, "ml_test.txt") + meta_path = os.path.join(dst_path, "ml_meta.txt") + users, items = set(), set() + history = {} + with open(file_path, 'r') as f: + lines = f.readlines() + for line in tqdm(lines): + user, item, score, timestamp = line.strip().split("::") + users.add(int(user)) + items.add(int(item)) + history.setdefault(int(user), []) + history[int(user)].append([item, timestamp]) + random.shuffle(list(users)) + with open(train_path, 'w') as f1, open(val_path, 'w') as f2, open(test_path, 'w') as f3: + for user in users: + hist = history[int(user)] + hist.sort(key=lambda x: x[1]) + for idx, value in enumerate(hist): + if idx == len(hist) - 1: + f3.write(str(user) + '\t' + value[0] + '\n') + elif idx == len(hist) - 2: + f2.write(str(user) + '\t' + value[0] + '\n') + else: + f1.write(str(user) + '\t' + value[0] + '\n') + with open(meta_path, 'w') as f: + f.write(str(max(users)) + '\t' + str(max(items))) + return train_path, val_path, test_path, meta_path +``` + +由于ml-1m数据集user、item的编码都是从1开始，因此无需重新编码。 + +当然在reclearn项目中大部分模型需要使用**用户的行为序列作为特征**，因此还提供了序列数据的划分方式。例如`data/datasets/games.py`中的`split_seq_data`方法： + +```python +def split_seq_data(file_path): + """split amazon games for sequence recommendation + Args: + :param file_path: A string. The file path of 'ratings_Beauty.dat'. + :return: train_path, val_path, test_path, meta_path + """ + dst_path = os.path.dirname(file_path) + train_path = os.path.join(dst_path, "games_seq_train.txt") + val_path = os.path.join(dst_path, "games_seq_val.txt") + test_path = os.path.join(dst_path, "games_seq_test.txt") + meta_path = os.path.join(dst_path, "games_seq_meta.txt") + users, items = set(), dict() + user_idx, item_idx = 1, 1 + history = {} + with open(file_path, 'r') as f: + lines = f.readlines() + for line in tqdm(lines): + user, item, score, timestamp = line.strip().split(",") + users.add(user) + if items.get(item) is None: + items[item] = str(item_idx) + item_idx += 1 + history.setdefault(user, []) + history[user].append([items[item], timestamp]) + with open(train_path, 'w') as f1, open(val_path, 'w') as f2, open(test_path, 'w') as f3: + for user in users: + hist_u = history[user] + if len(hist_u) < 4: + continue + hist_u.sort(key=lambda x: x[1]) + hist = [x[0] for x in hist_u] + time = [x[1] for x in hist_u] + f1.write(str(user_idx) + "\t" + ' '.join(hist[:-2]) + "\t" + ' '.join(time[:-2]) + '\n') + f2.write(str(user_idx) + "\t" + ' '.join(hist[:-2]) + "\t" + ' '.join(time[:-2]) + "\t" + hist[-2] + '\n') + f3.write(str(user_idx) + "\t" + ' '.join(hist[:-1]) + "\t" + ' '.join(time[:-1]) + "\t" + hist[-1] + '\n') + user_idx += 1 + with open(meta_path, 'w') as f: + f.write(str(user_idx - 1) + '\t' + str(item_idx - 1)) + return train_path, val_path, test_path, meta_path +``` + + + +## Criteo + +Criteo数据集大概有4500w的数据量，我们很难向处理moivelens那样将所有的数据都读如内存中。本文给出了两种处理方式： + +1. 只读取部分数据进行实验； +2. 将数据集进行切分若干份，模型训练时分别读取，这样就避免了难以读入的问题； + +**第一种方法**在`data/datasets/criteo.py`的`create_small_criteo_dataset`方法中： + +```python +def create_small_criteo_dataset(file, embed_dim=8, read_part=True, sample_num=100000, test_size=0.2): + """Load small criteo data(sample num) without splitting "train.txt". + Note: If you want to load all data in the memory, please set "read_part" to False. + Args: + :param file: A string. dataset's path. + :param embed_dim: A scalar. the embedding dimension of sparse features. + :param read_part: A boolean. whether to read part of it. + :param sample_num: A scalar. the number of instances if read_part is True. + :param test_size: A scalar(float). ratio of test dataset. + :return: feature columns such as [sparseFeature1, sparseFeature2, ...], + train, such as ({'C1': [...], 'C2': [...]]}, [1, 0, 1, ...]) + and test ({'C1': [...], 'C2': [...]]}, [1, 0, 1, ...]). + """ + if read_part: + data_df = pd.read_csv(file, sep='\t', iterator=True, header=None, + names=NAMES) + data_df = data_df.get_chunk(sample_num) + else: + data_df = pd.read_csv(file, sep='\t', header=None, names=NAMES) +``` + +通过`read_part`参数判断是否需要读取部分数据，`sample_num`为读取数据的总量。 + + + +**第二种方法**具体需要考虑的内容： + +1. 指定每个子数据集的数据量； +2. 子数据集存储的位置以及命名方式； +3. 如何实现数据集分割； + +通过`data/datasets/criteo.py`的`get_split_file_path`方法可以完成数据集的分割以及最终子数据集的相对存储路径获取： + +```python +def get_split_file_path(parent_path=None, dataset_path=None, sample_num=5000000): + """Get the list of split file path. + Note: Either parent_path or dataset_path must be valid. + If exists dataset_path + "/split", parent_path = dataset_path + "/split". + Args: + :param parent_path: A string. split file's parent path. + :param dataset_path: A string. + :param sample_num: A int. The sample number of every split file. + :return: A list. [file1_path, file2_path, ...] + """ + sub_dir_name = 'split' + if parent_path is None and dataset_path is None: + raise ValueError('Please give parent path or file path.') + if parent_path is None and os.path.exists(os.path.join(os.path.dirname(dataset_path), sub_dir_name)): + parent_path = os.path.join(os.path.dirname(dataset_path), sub_dir_name) + elif parent_path is None or not os.path.exists(parent_path): + splitByLineCount(dataset_path, sample_num, sub_dir_name) + parent_path = os.path.join(os.path.dirname(dataset_path), sub_dir_name) + split_file_name = os.listdir(parent_path) + split_file_name.sort() + split_file_list = [parent_path + "/" + file_name for file_name in split_file_name if file_name[-3:] == 'txt'] + return split_file_list +``` + +`parent_path`为存储子数据集的路径，`dataset_path`为原始数据集的路径，这里也判断了子数据集是否已经分割好，若之前已经完成，则可以直接获取子数据集的列表路径，避免重复的完成任务。若没有完成，我没通过`data/utils.py`的`splitByLineCount`方法完成： + +```python +def splitByLineCount(filename, count, sub_dir_name): + """Split File. + Note: You can specify how many rows of data each sub file contains. + Args: + :param filename: A string. + :param count: A scalar(int). + :param sub_dir_name: A string. + :return: + """ + f = open(filename, 'r') + try: + head = f.readline() + buf = [] + sub = 1 + for line in f: + buf.append(line) + if len(buf) == count: + sub = mkSubFile(buf, head, filename, sub_dir_name, sub) + buf = [] + if len(buf) != 0: + mkSubFile(buf, head, filename, sub_dir_name, sub) + finally: + f.close() +``` + +其中`mkSubFile`方法对当前达标目标说的内容`buf`进行存储，存储名字通过`sub`来指定。 + +```python +def mkSubFile(lines, head, srcName, sub_dir_name, sub): + """Write sub-data. + Args: + :param lines: A list. Several pieces of data. + :param head: A string. ['label', 'I1', 'I2', ...]. + :param srcName: A string. The name of data. + :param sub_dir_name: A string. + :param sub: A scalar(Int). Record the current number of sub file. + :return: sub + 1. + """ + root_path, file = os.path.split(srcName) + file_name, suffix = file.split('.') + split_file_name = file_name + "_" + str(sub).zfill(2) + "." + suffix + split_file = os.path.join(root_path, sub_dir_name, split_file_name) + if not os.path.exists(os.path.join(root_path, sub_dir_name)): + os.mkdir(os.path.join(root_path, sub_dir_name)) + print('make file: %s' % split_file) + f = open(split_file, 'w') + try: + f.writelines([head]) + f.writelines(lines) + return sub + 1 + finally: + f.close() +``` + +以上就完成了大数据集的切分，我们可以挑选任意一个子数据集作为测试集。 \ No newline at end of file diff --git a/docs/exeperiment.md b/docs/exeperiment.md new file mode 100644 index 0000000..542d4fc --- /dev/null +++ b/docs/exeperiment.md @@ -0,0 +1,18 @@ +# 实验说明 + +本篇将详细介绍如何构建简明的实验，实验案例可以参考[example](../example)。Reclearn的实验主要分为两大模块：**Top-K推荐**与**CTR 预估**，因此我们从这两部分内容进行展开。 + + + +## Top-K推荐 + +Top-K推荐指的是对于单一用户，推荐模型给其推荐物料库中最有可能的K个物品，然后再通过一些离线排序指标例如HR@K、NDCG@K、MRR@K、MAP@K等评估该模型的能力，在工业上这类模型也称为召回模型，应用在召回阶段。对于召回模型，**最显著的特点**就是user侧的user embedding向量与item embedding向量几乎不会发生特征交叉，除了最后阶段的点积。 + +Reclearn项目包含了学术上的模型（SASRec、AttRec等）和工业上的模型（DSSM、YoutubeDNN），但最终的实验流程需要统一，所以本篇实验说明定义了一种常用的数据集切割、负采样、离线指标评估等。如果使用者有自己的实验方式，也可以自己形式化定义。 + + + + + + +