SVM training

library('e1071')

frm6 <- read.csv("./frm6.csv")
frm7 <- read.csv("./frm7.csv")
fulldata <- merge(frm6, frm7, all=TRUE)

# pick only relevant data columns
fulldata <- fulldata[,c("radius", "angle", "class")]

# number of rows in the data frame
N <- nrow(fulldata)

# add new column of Bs
fulldata$fg <- rep("B",N)

# set to "F" those data samples which have "class > 1"
fulldata$fg[fulldata$class > 1] <- "F"

# leave only relevant data columns
fulldata <- fulldata[,c("fg","radius", "angle")]

# make FG column a column of factors
fulldata$fg <- factor(fulldata$fg)

Ntrain <- N * 0.7

fulldata.indices <- 1:N

# make an array of sample indices to use the selected samples for training
traindata.indices <- sample(fulldata.indices, Ntrain)

traindata <- fulldata[traindata.indices, ]
testdata <- fulldata[-traindata.indices, ]

#tuned <- tune.svm(fg ~., data=traindata, kernel="radial", gamma= 10^(-6:-1), cost = 10^(1:5))
#summary(tuned)

model <- svm(fg ~., data=traindata, kernel="radial", gamma=0.1, cost=1000)
print(model)

prediction <- predict(model, testdata[,c("radius", "angle")])
tbl <- table(pred = prediction, true = testdata$fg)

true_neg <- tbl[1,1]
true_pos <- tbl[2,2]

false_pos <- tbl[1,2]
false_neg <- tbl[2,1]
pos <- true_pos + false_pos
neg <- true_neg + false_neg

sensitivity <- true_pos / pos
specificity <- true_neg / neg
precision <- true_pos / (true_pos+false_pos)
recall <- true_pos / (true_pos + false_neg)
accuracy <- (true_pos + true_neg) / (pos + neg)

F1 <- 2 * (precision * recall) / (precision + recall)