Lab 1: Assignment 1 done

lab1/assignment1.R

@@ -3,7 +3,9 @@ library(dplyr)
 library(tidyr)
 library(kknn)
 library(ggplot2)
+library(DPpack)
 
+# 1. Import data
 data = read.csv("optdigits.csv", header=FALSE)
 
 n=dim(data)[1]
@@ -24,26 +26,32 @@ missclass=function(X,Xfit){
 }
 
 # Create model from training data
-model <- kknn(as.factor(V65)~., train, train, k=30, kernel="rectangular")
+model_train <- kknn(as.factor(V65)~., train, train, k=30, kernel="rectangular")
+model_test <- kknn(as.factor(V65)~., train, test, k=30, kernel="rectangular")
 
 # Get fitted values
-fitted <- model$fitted.values
+fitted_train <- model_train$fitted.values
+fitted_test <- model_test$fitted.values
 
 # Create confusion matrix
-confusion_matrix <- table(train$V65,fitted)
+confusion_matrix_train <- table(train$V65,fitted_train)
+confusion_matrix_test <- table(test$V65,fitted_test)
 
-print(confusion_matrix)
+print(confusion_matrix_train)
+print(confusion_matrix_test)
 
 # Get missclassification rate for the model
-train_missclassification <- missclass(train$V65,fitted)
+missclass_train <- missclass(train$V65,fitted_train)
+missclass_test <- missclass(test$V65,fitted_test)
 
-print(train_missclassification)
+print(missclass_train)
+print(missclass_test)
 
 # Get all cases where the target is 8
 digit_8_cases <- which(train$V65 == "8")
 
 # Get the probabilities for the cases
-probs_digit_8 <- model$prob[digit_8_cases, "8"]
+probs_digit_8 <- model_train$prob[digit_8_cases, "8"]
 
 easiest_cases <- digit_8_cases[order(-probs_digit_8)][1:2]
 hardest_cases <- digit_8_cases[order(probs_digit_8)][1:3]
@@ -80,8 +88,8 @@ print(train_missclassification)
 print(valid_missclassification)
 
 # Plot missclassification rates
-plot(1:30,valid_missclassification, ylim = c(0, max(valid_missclassification)), col="red", type="l")
-points(1:30, train_missclassification, col="blue", type="l")
+plot(1:30,valid_missclassification, ylim = c(0, max(valid_missclassification)), col="red")
+points(1:30, train_missclassification, col="blue")
 
 # Min classification rate for validation data
 print(which.min(valid_missclassification))
@@ -93,4 +101,26 @@ test_missclassification <- missclass(test$V65,model_test$fitted.values)
 # Missclassification rate for test data
 print(test_missclassification)
 
+cross_entropy=function(X_true,X_pred,epsilon=1e-15){
+
+  # Compute cross-entropy loss
+  return(-sum(X_true * log(X_pred + epsilon))/ nrow(X_true))
+}
+
+valid_cross_entropy <- numeric(30)
+
+for (i in 1:30) {
+  temp_model_valid <- kknn(as.factor(V65)~., train, valid, k=i, kernel="rectangular")
+  
+  
+  X_true <- model.matrix(~ as.factor(valid$V65) - 1)
+  X_pred <- temp_model_valid$prob
+
+  valid_cross_entropy[i] = cross_entropy(X_true,X_pred)
+}
+
+plot(1:30,valid_cross_entropy, ylim = c(0, max(valid_cross_entropy)), col="red", type="l")
+
+print(which.min(valid_cross_entropy))
+