Random Forest models.Rmd

---
title: "R Notebook"
output: html_notebook
---

```{r}
library(tidyverse)
library(caret)

tagged <- read_csv("Differing Descriptions Tagging (003).csv") %>%
  mutate(Change_importance = ifelse(Change_importance == 0, "Non.signif", "Signif")) %>%
  mutate(Addition_in_final = ifelse(is.na(Addition_in_final), "None", Addition_in_final),
         Deletion_from_draft = ifelse(is.na(Deletion_from_draft), "None", Deletion_from_draft))

set.seed(489)
train_partition <- createDataPartition(
  y = tagged$Change_importance,
  p = 0.8,
  list = F
)

train_data <- tagged[train_partition,]
test_data <- tagged[-train_partition,]

# ---------------------------
# Training PLS model
# ---------------------------

# Partial least squares discriminant analysis model, tuned over retained PLS components
pls_mod <- train(
  Change_importance ~ Draft_total + Final_total + match_ratio,
  data = train_data,
  method = "pls",
  preProc = c("center", "scale"), # Center/scale IVs for training
  trControl = trainControl(method = "repeatedcv", repeats = 3, classProbs = T, summaryFunction = twoClassSummary), # cross-fold validation instead of bootstrapping; use 10 folds, repeated 3 times
  tuneLength = 15,
  metric = "ROC"
)

pls_mod$results

ggplot(pls_mod) # Optimal ROC with 2 components

# ---------------------------
# Testing PLS model
# ---------------------------

pls_test_mod <- predict(pls_mod, newdata = test_data)
pls_preds <- predict(pls_mod, newdata = test_data, type = "prob")

confusionMatrix(data = pls_test_mod, as.factor(test_data$Change_importance)) # Predicts with 80% accuracy
```