Models_RF_LG_KNN.Rmd

---
title: "ADS-503 Team 4 Final Project"
author:
  - Leonard Littleton
  - Lina Nguyen
  - Emanuel Lucban
date: "6/13/21"
output:
  pdf_document: default
  html_notebook: default
  html_document:
    df_print: paged
---
```{r}
library(readr)
library(data.table)
library(mlbench)
library(ggplot2)
library(tidyr)
library(corrplot)
library(e1071)
library(caret)
library(naniar)
library(MLmetrics)
library(dplyr)
library(MASS)
library(pROC)
library(randomForest)
# Parallel Processing
library(doParallel)
cl <- makePSOCKcluster(4)
```


# Synthetic Financial Datasets For Fraud Detection

## Data Preparation

```{r}
synth_df <- fread('./data/PS_20174392719_1491204439457_log.csv', header = TRUE)
synth_df <- subset(synth_df, select = -c(isFlaggedFraud, nameOrig, nameDest))
synth_df <- cbind(synth_df, data.frame(hours_intraday = synth_df$step %% 24, 
                                           by_day = round(synth_df$step / 24), 
                                           day_of_week = round(synth_df$step / 24) %% 7))
# Log transform  (amount, oldbalanceOrg, newbalanceOrig, oldbalanceDest, newbalanceDest)
cont_vars <- c('amount', 'oldbalanceOrg', 'newbalanceOrig', 'oldbalanceDest', 'newbalanceDest')
# add small constant to prevent inf values
log_scaled <- sapply(data.frame(synth_df)[, cont_vars], function(x) log(x + 1))
colnames(log_scaled) <- lapply(cont_vars, function(x) paste('log_', x, sep=''))
synth_df <- cbind(synth_df, log_scaled)
synth_df$type <- as.factor(synth_df$type)
dmy <- dummyVars(" ~ type", data = synth_df, sep = '.', fullRank = TRUE)
synth_df <- cbind(synth_df, data.frame(predict(dmy, newdata = synth_df)))
# drop type
synth_df <- subset(synth_df, select = -c(type))
```

**Split Data into Training and Test Datasets using Stratified Random Sampling**
```{r}
set.seed(42)
# split x and y
x <- subset(synth_df, select = -c(isFraud))
y <- synth_df$isFraud
data_part <- createDataPartition(y = y, p = 0.75, list = FALSE)
x_train <- x[data_part, ]
y_train <- y[data_part]
x_test <- x[-data_part, ]
y_test <- y[-data_part]
y_train <- as.factor(y_train)
y_test <- as.factor(y_test)
levels(y_train) <- c('no', 'yes')
levels(y_test) <- c('no', 'yes')
```



##Random Forest
```{r}
#registerDoParallel(cl)
control <- trainControl(method = 'cv', number = 5, classProbs = TRUE, summaryFunction = prSummary)

gbm <- train(x_train, y_train, method = 'gbm', metric = 'AUC', 
            preProcess = c('center', 'scale'), 
            trControl = control)
saveRDS(lr, file = 'rds_files/gbm_model.rds')
```

```{r}
print(rf)
print(rf$finalModel)

#confusion matrix
set.seed(3)
rfpredict <- predict(rf, newdata = test)
confusionMatrix(rfpredict, test$y_test)

#AUC curve
rfroc <- pROC::roc(response = train$y_test, predictor = predictions[,1])
rfauc <- rfroc$auc[1]
rfauc
```
##knn
```{r}
registerDoParallel(cl)
control <- trainControl(method = 'cv', number = 5, classProbs = TRUE, summaryFunction = prSummary)
knn <- train(x_train, y_train, method = 'knn', preProcess = c('center', 'scale'), tuneLength = 5, trainControl = control, metric = 'AUC')
saveRDS(knn, file = 'rds_files/knn_model.rds')
```

```{r}
#confusion matrix
set.seed(3)
knnpredict <- predict(knn, newdata = x_test)
confusionMatrix(knnpredict, y_test)

##AUC curve
knnpredict <- predict(knn, x_train, type = 'prob')
knnroc <- pROC::roc(reponse = y_train, predictor = knnpredict[,1])
knnauc <- knnroc$auc[1]
knnauc
```

##Logistic Regression
```{r}

registerDoParallel(cl)
#registerDoSEQ()
ctrl <- trainControl(summaryFunction = prSummary, classProbs = TRUE)
lr <- train(x_train, y= y_train, method = 'glm', preProcess = c('center', 'scale'), metric = 'AUC', trControl = ctrl)
lr
saveRDS(lr, file = 'rds_files/lr_model.rds')
```
```{r}
lrModel <- readRDS('rds_files/lr_model.rds')
```

```{r}
lrPred <- predict(lrModel, newdata = x_test)
lrCFM <- confusionMatrix(lrPred, y_test, positive = 'yes')
lrProbs <- predict(lrModel, newdata = x_test, type = 'prob')
lrCFM
```

```{r}
confusionMatrix(data = lr$pred$pred, reference = lr$pred$obs)
lrROC <- roc(response = lr$pred$obs, predictor = lr$pred$successful, levels = rev(levels(lr$pred$obs)))
plot(lrROC, legacy.axes = TRUE)
auc(lrROC)


```


```{r}
lrROC <- roc(y_test, predict(lrModel, newdata = x_test, type = 'prob')$yes)
saveRDS(lrROC, file = 'rds_files/lr_roc.rds')

```