Project_backorder.Rmd

---
title: "intern"
author: "KEERTHI.R"
date: "January 31, 2018"
output: html_document
---

```{r setup, include=FALSE}
#knitr::opts_chunk$set(echo = TRUE)
```

```{r}
rm(list = ls())
#install.packages("")
library(ggplot2)
library(GGally)
library(moments)
library(car)
library(DMwR)
library(caret)
library(MASS)
library(glmnet)
library(doParallel)
library(rpart)
library(randomForest)
library(xgboost)
```





```{r}
train <- read.csv("train.csv",header = T,stringsAsFactors = F,na.strings = c("NA",""," "))
test <- read.csv("test.csv",header = T,stringsAsFactors = F,na.strings = c("NA",""," "))
```

#finding rows with na
```{r}
df = train[is.na(train$went_on_backorder),]
sum(is.na(df))
train = train[-c(	1687861),]
rm(df)

df = test[is.na(test$went_on_backorder),]
test = test[-c(242076),]
rm(df)
```
```{r}

```

```{r}

```


#PCA
```{r}
pcamodel <- preProcess(train_data[,!names(train_data)%in% c("went_on_backorder")],method="pca",pcaComp=3)

train_pca=predict(pcamodel,train_data)
test_pca=predict(pcamodel,test_data)



colSums(is.na(train_pca))      
colSums(is.na(test_pca)) 

train_pca<-knnImputation(train_pca[,!names(train_pca)%in% c("went_on_backorder")],k=5,scale=T)



train_pca= na.omit(train_pca)

```

```{r}
df$type = NULL
df$sku = NULL

```

```{r}
dim(df)
df$went_on_backorder <- as.factor(as.character(df$went_on_backorder))

str(df)

```



```{r}
for(i in 1:length(colnames(df))-1)  
{
  if(class(df[,i ]) == "character")  
  {
    df[,i] <- as.numeric(as.factor(df[,i]))
  }
}
```

```{r}
df$sku <- NULL
df_imputed <- centralImputation(df)

colSums(is.na(df_imputed))

```



```{r}
#str(df_imputed)
```




```{r}
#library(plyr)
#count(df_imputed$went_on_backorder)
```

#to find class imbalance
```{r}
prop.table(table(df_imputed$went_on_backorder))

barplot(prop.table(table(df_imputed$went_on_backorder)),
        col = rainbow(2),
        ylim = c(0,1),
        main = "Class Distribution")
```

```{r}
train_model <- createDataPartition(df_imputed$went_on_backorder, p=0.7, list = F)

train_data <- df[train_model,]

test_data <- df[-train_model,]


#rm(train)

```
#over sampaling
```{r}
library(ROSE)
under <- ovun.sample(went_on_backorder~., data = train_data, method = "under")$data
table(train_data$went_on_backorder)
table(under$went_on_backorder)
```


```{r}
str(df_imputed)

```

```{r}
specificity_train <- cm_train[1,1]/sum(cm_train[1, ])
```

```{r}
sensitivity_train <- cm_train[2,2]/sum(cm_train[2, ])
```


```{r}
pred.rf1_val <- predict(rf1,test_data)

cm_test <- table("actual" = test_data$went_on_backorder, "predicted" = pred.rf1_val)

specificity_test <- cm_test[1,1]/sum(cm_test[1, ])

sensitivity_test <- cm_test[2,2]/sum(cm_test[2, ])

```
```{r}

```


#logistic
```{r}
model1 <- glm(went_on_backorder~., data = train, family = binomial)
summary(model1)
```

```{r}
library(ROCR)
prob_train <- predict(model1, type = "response")

pred <- prediction(prob_train, over$went_on_backorder)
pref <- performance(pred, measure = "tpr", x.measure = "fpr")

```


```{r}
plot(pref, col= rainbow(10), colorize=T, print.cutoffs.at=seq(0,1,0.1))
```


```{r}
#memory.limit(4000)
prob_test <- predict(model1, over, type = "response")
preds_test <- ifelse(prob_test > 0.8 , "YES", "NO")

#actual_test_val <- ifelse(over$went_on_backorder=="0.5","YES","NO")
actual_test_val <- over$went_on_backorder

#confusionMatrix(preds_test, over$went_on_backorder, positive = "YES")
conf_matrix <- table(actual_test_val,preds_test)

print(conf_matrix)
```


#random forest
```{r}
rf1 <- randomForest(went_on_backorder~.,data=df_imputed,keep.forest= TRUE, ntree = 50)
#sum(is.na(train_data))
#train_data <- na.omit(train_data)
## Check the error on the validation data
pred.rf1_val <- predict(rf1,train_data)
print(rf1)
```

```{r}
pred_train <- predict()

cm_train <- table("actual" = train_data$went_on_backorder, "predicted" = pred.rf1_val)
```



```{r}


```

#randomforest
```{r}
set.seed(123)
rows <- createDataPartition(df_imputed$went_on_backorder,p=0.7,list = F)
train_data <- df_imputed[rows,]
val_data <- df_imputed[-rows,]


rf1 <- randomForest(went_on_backorder~.,data=train_data,keep.forest= TRUE, ntree = 50)
pred.rf1_val <- predict(rf1,train_data)

cm_train <- table("actual" = train_data$went_on_backorder, "predicted" = pred.rf1_val)


specificity_train <- cm_train[1,1]/sum(cm_train[1, ])

sensitivity_train <- cm_train[2,2]/sum(cm_train[2, ])

```



```{r}
pred.rf1_val1 <- predict(rf1,val_data)

cm_test <- table("actual" = val_data$went_on_backorder, "predicted" = pred.rf1_val1)

specificity_test <- cm_test[1,1]/sum(cm_test[1, ])

sensitivity_test <- cm_test[2,2]/sum(cm_test[2, ])
```

```{r}
sum(is.na(df))
library(imputeMissings)

```

```{r}
library(C50)

#Tree based model
c5_tree <- C5.0(went_on_backorder ~ . , train_data)

C5imp(c5_tree, metric = "usage")

c5_rules <- C5.0(went_on_backorder ~ . , train_data, rules = T)
summary(c5_rules)
```

```{r}
preds <- predict(c5_tree, val_data)
library(caret)

confusionMatrix(preds, val_data$went_on_backorder)
```