SVM

# reading data and loading libraries
setwd("~/R/wd/R_digits")
data_all=read.csv("train.csv")
library(parallel)
library(e1071)
library(caret)
library(ROCR)



# pre-processing

  # binary : 8 --> TRUE, others --> FALSE
  labels <- data_all[,1] == 8

  ## multiclass
  # labels <- as.factor(data_all[,1])

  # data.frame
  data_8 <- cbind(labels,data_all[,-1])

  # quantity (%) of observations to use
  q_data <- 0.1
  nb_pics <- nrow(data_8)
  idx <- sample(1:nb_pics, floor(nb_pics*q_data), replace = FALSE)
  data_8 <- data_8[idx,]

  #test/train full index
  nb_pics <- nrow(data_8)
  idx <- sample(1:nb_pics, floor(nb_pics*0.8), replace = FALSE)

  # test/train full sets
  training <- data_8[idx,]
  testing <- data_8[-idx,]

  # deleting empty cols
  del_zeros <- function(matrice) {which(apply(matrice,2,max) !=0)}
  nozeros <- del_zeros(training)
  training <- training[,nozeros]



# svm

  # starting time
  t <- Sys.time()

  model <- svm(labels ~ ., data = training, probability = TRUE, type = "C-classification",cost = 2, gamma = 1.4 * (1/(dim(training)[2]-1)))

  # total time
  t <- Sys.time() - t
  t


# analysis

  # predict
  pred <- predict(model, testing[,-1], probability = TRUE)

  # confusion matrix
    CM <- confusionMatrix(pred, testing[,1])
    acc <- CM$overall[1]
    conf <- CM$table
    conf
    acc

  # ROC curve using ROCR
  plot(col = "red", performance(prediction(attr(pred, "probabilities")[,2], testing[,1]), 'tpr', 'fpr'))