Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

submission Zhongyuan Zhang #135

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

submission Zhongyuan Zhang #135

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
173 changes: 152 additions & 21 deletions Assignment7.Rmd
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,71 +5,202 @@ date: "11/30/2016"
output: html_document
---

In the following assignment you will be looking at data from an one level of an online geography tutoring system used by 5th grade students. The game involves a pre-test of geography knowledge (pre.test), a series of assignments for which you have the average score (av.assignment.score), the number of messages sent by each student to other students about the assignments (messages), the number of forum posts students posted asking questions about the assignment (forum.posts), a post test at the end of the level (post.test) and whether or not the system allowed the students to go on to the next level (level.up).
In the following assignment you will be looking at data from an one level of an online geography tutoring system used by 5th grade students.
1.The game involves a pre-test of geography knowledge (pre.test)
2.A series of assignments for which you have the average score (av.assignment.score)
3.The number of messages sent by each student to other students about the assignments (messages)
4.The number of forum posts students posted asking questions about the assignment (forum.posts)
5.A post test at the end of the level (post.test)
6.Whether or not the system allowed the students to go on to the next level (level.up)

## Part I

#Upload data
#Get ready with the pacakages
```{r}
library(dplyr)
library(tidyr)
library(ggplot2)
library(rpart)
library(corrplot)
```

#Upload data
```{r}
df<-read.csv("online.data.csv",stringsAsFactors =FALSE, header = TRUE )
```

#Visualization
```{r}
#Start by creating histograms of the distributions for all variables (#HINT: look up "facet" in the ggplot documentation)
df1<-gather(df,Vars,Value,-id,-level.up)

#Then visualize the relationships between variables
#Distribution for all the continuous variable
ggplot(df1,aes(x=Value))+
geom_histogram()+
facet_grid(.~Vars,scales = "free")

#Try to capture an intution about the data and the relationships

#Plot for all the discreet variable
df1$level_up<-"Level_up"
ggplot(df1,aes(x=level_up,fill=factor(level.up)))+
geom_bar(position = "stack",width=0.2)+
scale_color_discrete("Level_up_option")

#Then visualize the relationships between variables
COR<-cor(df[,-c(1,7)])

corrplot(COR, order="AOE", method="circle", tl.pos="ld", type="lower",
tl.col="blue", tl.cex=0.8, tl.srt=20,
addCoef.col="black", addCoefasPercent = TRUE,
sig.level=0.50, insig = "blank")

#Try to capture an intuition about the data and the relationships
#1. All the numeric variables are positively correlated
#2. post.test.score and message(student sent to others to ask questions) are magnitudinally correlated with r scoring .94
#3. post.test.score and average assignment score are moderately to highly correlated as well with r scoring .76
#4. message(student sent to others to ask questions) and average assignment score are moderately to highly correlated with r scoring .71
#5. Number of posts in forum shows weak correaltion with other variables
```
#Classification tree
```{r}
#Create a classification tree that predicts whether a student "levels up" in the online course using three variables of your choice (As we did last time, set all controls to their minimums)

#Try to use two model to see which one is more accurate

#First one with these three variables:"post.test.score","messages","av.assignment.score" based on the correlation plot
#Second one with these three variables: "pre.test.score","post.test.score","messages" based on common sense

#pre_prune<-rpart.control(maxdepth = ,minsplit = )

training_sample<-sample(nrow(df),nrow(df)*0.75)
df_training<-df[training_sample,]
df_testing<-df[-training_sample,]


set.seed(726)
c.tree1<-rpart(level.up~post.test.score+messages+av.assignment.score,data=df_training,method="class")
c.tree2<-rpart(level.up~pre.test.score+post.test.score+messages,data=df_training,method="class")

#Plot and generate a CP table for your tree
plotcp(c.tree1)
plotcp(c.tree2)

printcp(c.tree1)
printcp(c.tree2)
#Generate a probability value that represents the probability that a student levels up based your classification tree
df_testing$predicted1<-predict(c.tree1,df_testing,type="prob")[,2]
df_testing$predicted2<-predict(c.tree2,df_testing,type="prob")[,2] # in order to do draw the ROC AND AUC

D1$pred <- predict(rp, type = "prob")[,2]#Last class we used type = "class" which predicted the classification for us, this time we are using type = "prob" to see the probability that our classififcation is based on.
#Last class we used type = "class" which predicted the classification for us, this time we are using type = "prob" to see the probability that our classififcation is based on.
```

## Part II
#Now you can generate the ROC curve for your model. You will need to install the package ROCR to do this.
```{r}
library(ROCR)
library(pROC)#(I am using a different package)
#Plot the curve(ROCR)
pred.detail1 <- prediction(df_testing$predicted1,df_testing$level.up)
plot(performance(pred.detail1, "tpr", "fpr"))+abline(0, 1, lty = 2)

pred.detail2 <- prediction(df_testing$predicted2,df_testing$level.up)
plot(performance(pred.detail2, "tpr", "fpr"))+abline(0, 1, lty = 2)

#Plot the curve(pROC)
ROC1<-roc(df_testing$level.up,df_testing$predicted1)
ROC2<-roc(df_testing$level.up,df_testing$predicted2)

plot(ROC1,col="red")
plot(ROC2,col="blue")

roc(df_testing$level.up,df_testing$predicted1,plot = TRUE,legacy.axes=TRUE,col="red",xlab="False Positive Percentage",ylab="True Positive Percentage",lwd=4)

#Plot the curve
pred.detail <- prediction(D1$pred, D1$level.up)
plot(performance(pred.detail, "tpr", "fpr"))
abline(0, 1, lty = 2)
roc(df_testing$level.up,df_testing$predicted2,plot = TRUE,legacy.axes=TRUE,col="blue",xlab="False Positive Percentage",ylab="True Positive Percentage",lwd=4)

#Calculate the Area Under the Curve
unlist(slot(performance(Pred2,"auc"), "y.values"))#Unlist liberates the AUC value from the "performance" object created by ROCR
#Here remember, if legacy.axes=TRUE wsa not specified, the x-axis would be specificty but not 1-specificty(False positive rate)

#Calculate the Area Under the Curve (ROCR)
unlist(slot(performance(pred.detail1,"auc"), "y.values")) #1
unlist(slot(performance(pred.detail2,"auc"), "y.values")) #0.8276047
#Unlist liberates the AUC value from the "performance" object created by ROCR

#Calculate the Area Under the Curve (pROC)
auc(ROC1) # 1
auc(ROC2) # 0.8276
#Now repeat this process, but using the variables you did not use for the previous model and compare the plots & results of your two models. Which one do you think was the better model? Why?
```
## Part III
#Thresholds
```{r}
#Look at the ROC plot for your first model. Based on this plot choose a probability threshold that balances capturing the most correct predictions against false positives. Then generate a new variable in your data set that classifies each student according to your chosen threshold.
# Look at the ROC plot for your first model. Based on this plot choose a probability threshold that balances capturing the most correct predictions against false positives. Then generate a new variable in your data set that classifies each student according to your chosen threshold.

threshold.pred1 <-
# Since tree1 is too perfect to predict,I decide to observe tree 2's ROC plot
#Find the threshold
roc.info1<-roc(df_testing$level.up,df_testing$predicted1,plot = TRUE,legacy.axes=TRUE)
roc.info2<-roc(df_testing$level.up,df_testing$predicted2,plot = TRUE,legacy.axes=TRUE)

#Now generate three diagnostics:
roc.df1<-data.frame(tpp=roc.info1$sensitivities*100,
fpp=(1-roc.info1$specificities)*100,
thresholds=roc.info1$thresholds)
head(roc.df1)

D1$accuracy.model1 <-
roc.df2<-data.frame(tpp=roc.info2$sensitivities*100,
fpp=(1-roc.info2$specificities)*100,
thresholds=roc.info2$thresholds)
head(roc.df2) # find the optimal threshoulds:0.4078812

D1$precision.model1 <-
# since we have already given up the first
```

D1$recall.model1 <-
```{r}
# using the new threshold to classify
df_testing$threshold.pred2 <-ifelse(df_testing$predicted2>0.4078812,"yes","no")

#Finally, calculate Kappa for your model according to:
# Now generate three diagnostics:
# Creating a confusion_matrix
confusion_matrix<-as.matrix(table(df_testing$threshold.pred2,df_testing$level.up))

#First generate the table of comparisons
table1 <- table(D1$level.up, D1$threshold.pred1)

#create a function to calculate Accuracy, Precision and Recall(for 2 by 2 matrix only)
diagnositics_cal<-function(type="a",matrix){
TN<-matrix[1]
FN<-matrix[3]
FP<-matrix[2]
TP<-matrix[4]
if(type=="p"){
out<-TP/(TP+FN)
}else if(type=="r"){
out<-TP/(TP+FP)
}else if(type=="a"){
out<-(TP+TN)/(TP+TN+FP+FN)
}else{
print("you need to specify what type of calculation you are using")
}
out
}


df_testing$accuracy.model2<-diagnositics_cal("a",confusion_matrix)
df_testing$precision.model2 <- diagnositics_cal("p",confusion_matrix)
df_testing$recall.model2 <-diagnositics_cal("r",confusion_matrix)

#model2 accuracy is 0.808, precision is 0.9183673, recall is 0.6923077


#Finally, calculate Kappa for your model according to:
#First generate the table of comparisons
table(df_testing$threshold.pred2,df_testing$level.up)
#Convert to matrix
matrix1 <- as.matrix(table1)
matrix1 <- as.matrix(table(df_testing$threshold.pred2,df_testing$level.up))


#Calculate kappa
kappa(matrix1, exact = TRUE)/kappa(matrix1)

# kappa is not possible >1
#install.packages("psych")
library(psych)
cohen.kappa(matrix1)

#Calculate kappa
kappa(matrix1, exact = TRUE)/kappa(matrix1)
Expand Down
Loading