Lecture 9 Spring 2016-Comparing ML models

# Load Packages
library(mlbench)

## Warning: package 'mlbench' was built under R version 4.5.2

library(caret)

## Warning: package 'caret' was built under R version 4.5.2

## Loading required package: ggplot2

## Warning: package 'ggplot2' was built under R version 4.5.2

## Loading required package: lattice

data("PimaIndiansDiabetes")

#Prepare training scheme

trainControl <- trainControl(method = "repeatedcv", number = 10, repeats = 3)

#CART
set.seed(7)
fit.cart <- train(diabetes~.,data=PimaIndiansDiabetes, method="rpart", trControl=trainControl)

# LDA
 set.seed(7)
 fit.lda <- train(diabetes~., data=PimaIndiansDiabetes, method="lda", trControl=trainControl)
 
# SVM
 set.seed(7)
fit.svm <- train(diabetes~., data=PimaIndiansDiabetes, method="svmRadial", trControl=trainControl)
# KNN
set.seed(7)
fit.knn <- train(diabetes~., data=PimaIndiansDiabetes, method="knn", trControl=trainControl)

# Random Forest
 set.seed(7)
 fit.rf <- train(diabetes~., data=PimaIndiansDiabetes, method="rf", trControl=trainControl)

#Collect resamples

results<-resamples(list(CART=fit.cart, LDA=fit.lda, SVM=fit.svm, KNN=fit.knn, RF=fit.rf))

#Summarize differences between models

summary(results)

## 
## Call:
## summary.resamples(object = results)
## 
## Models: CART, LDA, SVM, KNN, RF 
## Number of resamples: 30 
## 
## Accuracy 
##           Min.   1st Qu.    Median      Mean   3rd Qu.      Max. NA's
## CART 0.6753247 0.7272727 0.7532468 0.7469697 0.7662338 0.7922078    0
## LDA  0.7142857 0.7508117 0.7662338 0.7791069 0.8000256 0.9078947    0
## SVM  0.7236842 0.7508117 0.7631579 0.7712919 0.7915243 0.8947368    0
## KNN  0.6753247 0.7036056 0.7272727 0.7369503 0.7662338 0.8311688    0
## RF   0.6842105 0.7305195 0.7597403 0.7638528 0.8019481 0.8421053    0
## 
## Kappa 
##           Min.   1st Qu.    Median      Mean   3rd Qu.      Max. NA's
## CART 0.2762566 0.3620724 0.4241878 0.4151867 0.4861107 0.5250000    0
## LDA  0.3011551 0.4192537 0.4662541 0.4862025 0.5308596 0.7812500    0
## SVM  0.3391908 0.3997116 0.4460612 0.4621585 0.5234605 0.7475083    0
## KNN  0.2553191 0.3406000 0.3841761 0.3984995 0.4539789 0.6195363    0
## RF   0.2951613 0.3778304 0.4640696 0.4630809 0.5447483 0.6426332    0

#box and whisker plots to compare models

scales <- list(x=list(relation="free"),y=list(relation="free"))
bwplot(results, scales=scales)

#density plots of accuracy

scales <- list(x =list(relation="free", y =list(relation="free")))
densityplot(results, scales=scales, pch = "|")

## Warning in complete_names(x, x.scales): Invalid or ambiguous component names: y

#pairwise scatter plots of predictions to compare models

splom(results)

#Calculate and summarize statistical significance
#difference in model predictions
diffs<-diff(results)

#summarize p-values for pairwise comparisons
summary(diffs)

## 
## Call:
## summary.diff.resamples(object = diffs)
## 
## p-value adjustment: bonferroni 
## Upper diagonal: estimates of the difference
## Lower diagonal: p-value for H0: difference = 0
## 
## Accuracy 
##      CART      LDA       SVM       KNN       RF       
## CART           -0.032137 -0.024322  0.010019 -0.016883
## LDA  0.0011862            0.007815  0.042157  0.015254
## SVM  0.0116401 0.9156892            0.034342  0.007439
## KNN  1.0000000 6.68e-05  0.0002941           -0.026902
## RF   0.2727542 0.4490617 1.0000000 0.0183793          
## 
## Kappa 
##      CART      LDA        SVM        KNN        RF        
## CART           -0.0710158 -0.0469717  0.0166872 -0.0478942
## LDA  0.0008086             0.0240440  0.0877029  0.0231215
## SVM  0.0258079 0.3562734              0.0636589 -0.0009225
## KNN  1.0000000 0.0003858  0.0040823             -0.0645814
## RF   0.0211763 1.0000000  1.0000000  0.0158974