project.Rmd

---
title: "IE5374Project"
author: "Dhaval Jariwala"
date: "2022-12-06"
output: html_document
---
## IE5374 PROJECT

## BACKORDER PREDICTION


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

library(ggplot2)
library(tidyverse)
library(dplyr)
library(ROSE)
library(ROCR)
library(caret)
library(MASS)
library(gridExtra)

trainData <- read.csv("BackOrder_train.csv", stringsAsFactors = T)[,-1]
testData <- read.csv("BackOrder_test.csv", stringsAsFactors = T)[,-1]
data <- rbind(trainData,testData)
```

## CHECK THE DATASETS
```{r}
#TRAINING DATA

head(trainData)

# Check dimensions of the dataset
dim(trainData)

# check the datatype of each column
sapply(trainData,class)

# summary of all columns in the dataset
summary(trainData)
```

```{r}
# TESTING DATA

head(testData)

# Check dimensions of the dataset
dim(testData)

# check the datatype of each column
sapply(testData,class)

# summary of all columns in the dataset
summary(testData)

```

## EDA
```{r}
ggplot(data = trainData, aes(x = sales_3_month, y = forecast_3_month)) + geom_point(aes(color = factor(went_on_backorder))) 

ggplot(data = trainData+ aes(x = sales_6_month, y = forecast_6_month)) + geom_point(aes(color = factor(went_on_backorder))) 

ggplot(data = trainData ,aes(x = sales_9_month, y = forecast_9_month)) + geom_point(aes(color = factor(went_on_backorder))) 

ggplot(data = trainData, aes(x = min_bank, y = forecast_3_month)) + geom_point(aes(color = factor(went_on_backorder)))

ggplot(data = trainData, aes(x = min_bank, y = forecast_6_month)) + geom_point(aes(color = factor(went_on_backorder))) 

ggplot(data = trainData, aes(x = min_bank, y = forecast_9_month)) + geom_point(aes(color = factor(went_on_backorder)))

ggplot(data=PrductsData, aes(x=went_on_backorder, y=national_inv, colour = lead_time, size = in_transit_qty)) + 
    geom_point() + 
    xlab("went_on_backorder")+ylab("national_inv") + 
    ggtitle("went_on_backorder Vs. national_inv")


```



## DATA CLEANING
```{r}
# TRAINING DATA

#checking the outcome variable for mislabled entries
rawTrainData = trainData
p1 = ggplot(rawTrainData, aes(went_on_backorder)) + geom_bar() + ggtitle("Before Cleaning")

# from the summary it is evident that there is one row which is incorrectly entered in the dataset, it has either missing or null values
trainData <- trainData %>% filter(went_on_backorder != "")

# remove the rows with all NA
trainData <- trainData[apply(trainData, 1, function(y) !all(is.na(y))),]

# remove the Cols with all NA
trainData <- trainData[sapply(trainData, function(x) !all(is.na(x)))]
  
# drop unused levels induced by null row
trainData <- droplevels(trainData)

dim(trainData)

summary(trainData)

p2 = ggplot(trainData, aes(went_on_backorder)) + geom_bar() + ggtitle("After Cleaning")
grid.arrange(p1, p2, ncol = 2)  

# rows and cols with all NA values and un-used levels are removed from the data set
```

```{r}
# TESTING DATA

#checking the outcome variable for mislabled entries
rawTestData = testData
p1 = ggplot(rawTestData, aes(went_on_backorder)) + geom_bar() + ggtitle("Before Cleaning")

# from the summary it is evident that there is one row which is incorrectly entered in the dataset, it has either missing or null values
testData <- testData %>% filter(went_on_backorder != "")

# remove the rows with all NA
testData <- testData[apply(testData, 1, function(y) !all(is.na(y))),]

# remove the Cols with all NA
testData <- testData[sapply(testData, function(x) !all(is.na(x)))]
  
# drop unused levels induced by null row
testData <- droplevels(testData)

dim(testData)

summary(testData)

p2 = ggplot(testData, aes(went_on_backorder)) + geom_bar() + ggtitle("After Cleaning")

grid.arrange(p1,p2, ncol = 2)

# rows and cols with all NA values and unsed levels are removed from the data set

```
## Outlier Detection

```{r}
## TRAINING DATA
continuousVariables <- select_if(trainData,is.numeric)
#continuousVariables <- as.data.frame(colnames(continuousVariables)) 
#colnames(continuousVariables) <- c("names")
continuousVariables <- colnames(continuousVariables)
continuousVariables

library(robustHD)

z = trainData
z[,continuousVariables] <- sapply(z[,continuousVariables], function(x) winsorize(x, threshold =1.5, method = "zscore", robust = TRUE))

#par(mfrow=c(1,2))
p1 = boxplot(trainData[,continuousVariables],  xaxt = "n")
tick <- seq_along(p1$names)
axis(1, at = tick, labels = FALSE)
text(tick, par("usr")[3] - 0.45, p1$names, srt = 45, xpd = TRUE)

p2 = boxplot(z[,continuousVariables], xaxt = "n")
tick <- seq_along(p2$names)
axis(1, at = tick, labels = FALSE)
text(tick, par("usr")[3] - 0.45, p2$names, srt = 45, xpd = TRUE)

c = cor(trainData[,continuousVariables])

library('corrplot')
corrplot(c, method="color")

```


## Data Imputation (solve missing values problem)
```{r}
# TRAINING DATA

a<-sapply(trainData,function(x)sum(is.na(x)))

data.frame(a,variable=colnames(trainData))%>%
  ggplot(aes(reorder(variable,-a),a))+
    geom_bar(stat="identity")+
    labs(x="variable")+
    theme(axis.text.x = element_text(angle=90))

# it is seen that the column "lead_time" has a lot of missing values
# therefore we replace it with mean value of the column lead_time

trainData$lead_time[is.na(trainData$lead_time)] = mean(trainData$lead_time, na.rm=TRUE)

summary(trainData)
```

```{r}
# TESTING DATA

b<-sapply(testData,function(x)sum(is.na(x)))

data.frame(a,variable=colnames(testData))%>%
  ggplot(aes(reorder(variable,-b),b))+
    geom_bar(stat="identity")+
    labs(x="variable")+
    theme(axis.text.x = element_text(angle=90))

# it is seen that the column "lead_time" has a lot of missing values
# therefore we replace it with mean value of the column lead_time

testData$lead_time[is.na(testData$lead_time)] = mean(testData$lead_time, na.rm=TRUE)

summary(testData)

```

## CONVERTING ALL CHARCHTER VALUES TO NUMERIC
```{r}
#function to map yes and no values to 1 and 0
Mapvalue <- function(x) {
  plyr::mapvalues(x, from = c("Yes", "No"), to = c(1,0))
}

# TRAINING DATA
categoricalData1<-select_if(trainData,is.factor)
categoricalData1[c(1:ncol(categoricalData1))] <- lapply(categoricalData1[c(1:ncol(categoricalData1))], Mapvalue)
NumericData1<-select_if(trainData,is.numeric)
trainData <- cbind(NumericData1, categoricalData1)
head(trainData)

# TESTING DATA
categoricalData2<-select_if(testData,is.factor)
categoricalData2[c(1:ncol(categoricalData2))] <- lapply(categoricalData2[c(1:ncol(categoricalData2))], Mapvalue)
NumericData2<-select_if(testData,is.numeric)
testData <- cbind(NumericData2, categoricalData2)
head(testData)


```





## SAMPLING
```{r}
# It is clearly seen that the outcome variable 'went_on_backorder' is biased, therefore it is necessary to oversample the data so as to get a better accuracy and fit on the model

sampledTrainData <- ovun.sample(went_on_backorder~., data = trainData, method = "both")$data
table(trainData$went_on_backorder)
table(sampledTrainData$went_on_backorder)

grid.arrange(
  ggplot(trainData, aes(went_on_backorder)) + geom_bar() , 
  ggplot(sampledTrainData, aes(went_on_backorder)) + geom_bar(), 
  ncol = 2) 

 
```
## MODELING

```{r}
# LOGISTIC REGRESSION
fit.glm1 <- caret::train(went_on_backorder~ lead_time +sales_1_month+ national_inv  + sales_3_month + forecast_9_month , data=trainData, method="glm",family= binomial)
varImp(fit.glm1)
fit.glm1$results

```

```{r}
fit.glm2 <- caret::train(went_on_backorder~ lead_time + national_inv , data=underSampledTrainData, method="glm",family= binomial)
varImp(fit.glm2)
fit.glm2$results
```

```{r}
# DECISION TREE
fit.dt1 <- caret::train(went_on_backorder~., data=underSampledTrainData, method="rpart")
varImp(fit.dt1)
fit.dt1$results
```

```{r}
fit.dt2 <- caret::train(went_on_backorder~ lead_time +sales_9_month+ national_inv  + sales_6_month + forecast_3_month , data=underSampledTrainData, method="rpart")
varImp(fit.dt2)
fit.dt2$results
```

## RESULTS

```{r}
results <- resamples(list(GLM=fit.glm1,CART=fit.dt2))
summary(results)
```

```{r}
summary(fit.glm1)
```

```{r}
caret::confusionMatrix(underSampledTrainData$went_on_backorder,predict(fit.glm1),positive = "1")
```

## Prediction
```{r}
test.glm <- predict(fit.glm1, newdata = testData)
Conflog<-confusionMatrix(data = test.glm , reference =testData$went_on_backorder ,positive = "1")
Conflog
```

```{r}
caret::confusionMatrix(underSampledTrainData$went_on_backorder,predict(fit.dt1),positive = "1")
```

```{r}
test.DT <- predict(fit.dt1, newdata = testData)
#-- Confusion matrix of test data for Decision tree
Confdt<-confusionMatrix(data = test.DT , reference =testData$went_on_backorder ,positive = "1")
Confdt
```
## Result of Predicted Model
```{r}
Results1<-cbind(Conflog$byClass,Confdt$byClass)
Results1
```