vinothrandomforest.Rmd

---
title : "SANTANDER CUSTOMER SATISFACTION"
author: "VINOTHKUMAR A"
date  : "03 NOV 2018"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```
#include library
```{r}
library(ggplot2) # Data visualization
library(readr) # CSV file I/O, e.g. the read_csv function
```
# Santander Customer Satisfaction
# Comparison of unbalanced and downscaled Random Forests
```{r}
library(dplyr)
library(caret) 
library(randomForest)
```
# Read data
```{r}
train0 <- read.csv("D:/santander/train.csv") 
test0 <- read.csv("D:/santander/train.csv")
```
# First combine all the data for cleaning
```{r}
all_data <- rbind(train0,test0)
```
# Find predictors that have very low percentage of unique values
# zeroVar: gives vector of logicals for whether the predictor has only one disticnt value
# nzv : gives vector of logicals for whether the predictor is a near zero variance predictor
```{r} 
nsv <- nearZeroVar(all_data[,-1], 
                   saveMetrics = TRUE) # Do not test ID
```

# Subset, remove near zero variance predictors 
# Note: Here we could have chosen zeroVar, which would lead to a larger set of predictors
```{r}
all_data.new <- all_data[, !nsv$nzv]
```
#Visualization
```{r}
library(lattice)
splom(train0[,3:5]) #only for integer values

```
#correlation plot #to find the amount of correlation between variables
```{r}
library(corrplot)
cr <- cor(train0[,5:7])
corrplot(cr, method = "number")

```

# Find duplicate columns and then remove them
```{r}
dpl.cl <- data.frame(i = integer(), j = integer())
for (i in 2:(ncol(all_data.new)-1) ){
  for 
  (j in (i+1):ncol(all_data.new) ){
    
    if (identical(all_data.new[,i], all_data.new[,j]) == TRUE){
      dpl.cl <- rbind(dpl.cl, data.frame(i = i, j=j))
      #print(c(i,j))
      
    }
  }
}
all_data.new <- all_data.new[, -dpl.cl$j]

```

# Now split back into training and testing sets
```{r}
train1 <- all_data.new[1:nrow(train0), ]; train1$TARGET <- train0$TARGET
test1 <- all_data.new[(nrow(train0) + 1):nrow(all_data.new), ]
rm(train0,test0) # Remove unnecessary data.frames
```
# Turn TARGET into factor (0 and 1)
```{r}
train1 <- mutate(train1, TARGET = factor(TARGET))
nmin = sum(train1$TARGET == "1") #skewed class is 1
nmax = sum(train1$TARGET == "0")
c(nmin, nmax) # Much more 1's than 0's
```

# Create training and test data from train1
```{r}
set.seed(101)
inTrain <- createDataPartition(y = train1$TARGET, p = 0.7, list = FALSE)
training <- train1[inTrain, ]; testing <- train1[-inTrain, ]
```

# Let us try to train a simple Random Forest using down-sampling
```{r}
nmin = sum(training$TARGET == "1")
```

## Tell randomForest to sample by strata. 
## Specify that the number of samples selected within each class should be the same
```{r}
rf0 <- randomForest(TARGET ~., data = training, ntree = 500, mtry = 10,
                    strata = training$TARGET, sampsize = rep(nmin,2))
pred.rf0 <- predict(rf0, newdata = testing)
cm0 <- confusionMatrix(pred.rf0, testing$TARGET)

```
# Let's look at the unbalanced RF
```{r}
rfu <- randomForest(TARGET ~. , data = training, ntree = 500, mtry = 10)
pred.rfu <- predict(rfu, newdata = testing)
cmu <- confusionMatrix(pred.rfu, testing$TARGET)
```
# Compare the confusion matrices cmu and cm0, you will see the difference
# Let's look at the ROC curves
```{r}
library(pROC)
pred.rf0 <- predict(rf0, newdata = testing, type = "prob")[,1]
pred.rfu <- predict(rfu, newdata = testing, type = "prob")[,1]
downsampledROC <- roc(response = testing$TARGET, predictor = pred.rf0,
                      levels = rev(levels(testing$TARGET)))
unbalancedROC <- roc(response = testing$TARGET, predictor = pred.rfu,
                     levels = rev(levels(testing$TARGET)))
plot(downsampledROC, col = rgb(1, 0, 0, .5), lwd = 2); 
plot(unbalancedROC, col = rgb(0, 0, 1, .5), lwd = 2, add = TRUE)
```
# Area under the ROC curves
```{r}
auc(downsampledROC)
auc(unbalancedROC)
```

```{r}
submission <-data.frame(Id = test1$id,TARGET=pred.rf0)
write.csv(submission,"D:santander/submit3.csv",row.names=F)
```

# kaggle score is 0.500000 using RandomForset