Testing out a few things

.Rapp.history

@@ -0,0 +1,74 @@
+version()
+help()
+quit()
+library(rj)
+quit()
+?ln
+?log
+1/0.05 * (10 * log(3) + log(1/0.05))
+quit()
+epsilon <- 0.05
+delta <- 0.05
+vch <- 56
+1/epsilon(4*log2(2/delta) + 8*vch*log2(13/epsilon))
+1/epsilon * (4*log2(2/delta) + 8*vch*log2(13/epsilon))
+vch
+delta
+epsilon
+quit()
+install.packages("R.matlab")
+library(R.matlab)
+?R.matlab
+quit()
+install.packages("neuralnet")
+library(neuralnet)
+getwd()
+ls()
+?multinom
+library(nnet)
+?multinom
+getwd()
+library(nnet)
+data(iris)
+set.seed(1234)
+x <- iris[sample(1:nrow(iris)), ]
+dim(x)
+x <- cbind(x, useless = rnorm(nrow(x)))
+?cbind
+x[, 1] <- x[, 1] + rnorm(nrow(x))
+x[, 2] <- x[, 2] + rnorm(nrow(x))
+x[, 3] <- x[, 3] + rnorm(nrow(x))
+plot(x, col = x$Species)
+names(x)
+head(x)
+table(x$Species)
+train <- x[1:100, ]
+test <- x[101:150, ]
+library(caret)
+install.packages("caret")
+library(caret)
+model.knn3 <- knn3(Species ~ ., k = 5, data = train)
+?knn3
+model.knn3
+summary(model.knn3)
+predict(model.knn3, test, type = "class")
+table(true = test$Species, predicted = predict(model.knn3, test, type = "class"))
+?nnet
+model.nnet <- nnet(Species ~ ., data = train, size = 10)
+model.nnet
+predict(model_nnet, test)
+predict(model.nnet, test, type = "class")
+predict(model.nnet, test)
+table(true = test$Species, predicted = predict(model.nnet, test, type = "class"))
+?nnet
+quit()
+setwd("./Documents/workspace/601-project-new/")
+list.files()
+install.packages("randomForest")
+library(randomForest)
+rfNews()
+?randomForest
+source("./code/runHCL-SVM.R")
+out.filename
+ls()
+quit()

.gitignore

@@ -0,0 +1,3 @@
+.Rproj.user
+.Rhistory
+.RData

code/runHCL-SVM.R

@@ -20,17 +20,23 @@ Y.train <- read.csv(file = Y.train.filename, header = FALSE)
 Y.factor <- as.factor(Y.train$V1)
 
 train.data <- as.matrix(X.train)
+test.data <- as.matrix(X.test)
 train.pca <- prcomp(train.data, scale. = TRUE, tol = 0)
+test.pca <- prcomp(test.data, scale. = TRUE, tol = 0)
 
-# Choose the number of PC's
-epsilon <- 0.015
-nPrinComps <- 20
-nPrinComps <- determineNumPrinComps(train.data, train.pca, epsilon, nPrinComps)
+# # Choose the number of PC's
+# epsilon <- 0.015
+# nPrinComps <- 20
+# nPrinComps <- determineNumPrinComps(train.data, train.pca, epsilon, nPrinComps)
+nPrinComps <- 500
 
-out.filename <- "./figures/agnes-train-ward.pdf"
+# out.filename <- "./figures/agnes-train-ward.pdf"
+out.filename <- ""
 K <- 4 # choose 4 clusters
 train.data.cl <- getAgnesClusters(train.data, train.pca, nPrinComps, out.filename, K)
 projTrainData <- train.data.cl[, 1:nPrinComps]
+projTestData <- test.pca$x # get the rotated rata
+projTestData <- projTestData[, 1:nPrinComps]
 
 # Plot the K clusters projected on the first 2 PC's
 out.filename <- paste("./figures/agnes-train-cl-", K, ".pdf", sep = "")
@@ -39,13 +45,8 @@ mainStr <- paste("Hierarchical Clustering with K =", K)
 plot(projTrainData, col = train.data.cl$Cluster, main = mainStr)
 dev.off()
 
-test.data <- as.matrix(X.test)
 # Use k-NN to predict the class of test data
-test.cl <- knn(train.data, test.data, cl = train.data.cl$Cluster, k = 5)
-
-test.pca <- prcomp(test.data, scale. = TRUE, tol = 0)
-projTestData <- test.pca$x # get the rotated rata
-projTestData <- projTestData[, 1:nPrinComps]
+test.cl <- knn(projTrainData, projTestData, cl = train.data.cl$Cluster, k = 5)
 
 # Plot the K clusters projected on the first 2 PC's
 out.filename <- paste("./figures/knn-test-cl-", K, ".pdf", sep = "")
@@ -54,11 +55,12 @@ mainStr <- paste("k-NN Clustering with K =", K)
 plot(projTestData, col = test.cl, main = mainStr)
 dev.off()
 
+new.train.data <- as.data.frame(train.data)
+new.train.data <- cbind(new.train.data, class = Y.factor)
 all.preds <- data.frame()
 for(k in 1:K) {
-  subset.train <- train.data[train.data.cl$Cluster == k, ]
-  subset.Y.factor <- Y.factor[train.data.cl$Cluster == k]
-  subset.svm <- svm(x = subset.train, y = subset.Y.factor, kernel = "radial")
+  subset.train <- new.train.data[train.data.cl$Cluster == k, ]
+  subset.svm <- svm(class ~ ., data = subset.train, kernel = "radial")
 
   subset.test <- test.data[test.cl == k, ]
   subset.indices <- which(test.cl == k)

code/runPCA-ANN.R

@@ -33,8 +33,8 @@ projData <- projData[, 1:nPrinComps]
 new.train.data <- projData[1:nrow(train.data), ]
 new.train.data <- as.data.frame(new.train.data)
 new.train.data <- cbind(new.train.data, class = Y.factor)
-model.ann <- nnet(class ~ ., data = new.train.data, size = 20,
-                  linout = TRUE, MaxNWts = 3000)
+model.ann <- nnet(class ~ ., data = new.train.data, size = 250,
+                  linout = TRUE, MaxNWts = 300000)
 
 new.test.data <- projData[(nrow(train.data)+1):nrow(projData), ]
 preds <- predict(model.ann, new.test.data)
@@ -43,5 +43,5 @@ preds <- predict(model.ann, new.test.data)
 cat.preds <- max.col(preds)
 pred.data <- data.frame(Id = 1:nrow(test.data), Category = cat.preds)
 
-out.filename <- "./data/ANN-PCA.csv"
+out.filename <- paste("./data/ANN-PCA-", nPrinComps, ".csv", sep = "")
 write.csv(pred.data, file = out.filename, row.names = FALSE)