# 1. Naive Bayes classificator
# 2. SVM
# 3. KNN

# Пакет caret является своего рода мета-пакетом, который предоставляет универсальный способ 
# работы со множеством пакетов машинного обучения.
# install.packages("caret")


library(caret)
setwd("C:/Alexeeva/sent/RGNG_koltcov_coments_data/")
set.seed(23)  # установили сид, чтобы можно было воспроизвести результаты


### data processing
data <- read.delim2("comments_sentiment_data_mtx.csv", 
                    encoding="UTF-8", header=T, stringsAsFactors=FALSE)

# View(subset(data, Meta2 == "#„…‹/0!"))  # 8 broken rows
# as.numeric(rownames((subset(data, Meta2 %in% c("#„…‹/0!","")))))  # their ids: 1352,2045,3839,3974,4271,4593,5206,7184

data2 <- data[-c(1352,2045,3839,3974,4271,4593,5206,7184),c(2:4247)]
data2$myclass <- as.numeric(data$Meta2[-c(1352,2045,3839,3974,4271,4593,5206,7184)])

data2[,1] <- as.numeric(data2[,1])
data2[,2] <- as.numeric(data2[,2])
data2[,3] <- as.numeric(data2[,3])

saveRDS(data2, "comments-processed.RDS")
rm(data, data2)

### load processed data
data <- readRDS("comments-processed.RDS")

data$myclass <- factor( round(data$myclass))
levels(data$myclass)

inTrain <- createDataPartition(y = data$myclass, p = .75, list = FALSE)  # делим датасет на тренировочный (75%) и тестируемый

training <- data[ inTrain,-4246]  # -4246 exclude column with txt ids
testing  <- data[-inTrain,-4246]


saveRDS(training, "./data-round/training.RDS")
saveRDS(testing, "./data-round/testing.RDS")

ctrl <- trainControl(method = "cv", number=10, sampling="up")  # настраиваем параметры валидации модели: 10-fold сross validation

# Naive Bayes
start.time <- Sys.time()
nb_fit <- train(myclass ~ .,
                data = training, 
                method = "nb",
                trControl = ctrl)
end.time <- Sys.time()
time.taken <- end.time - start.time
saveRDS(nb_fit, "./models-round/naive-bayes.RDS")

# тестируем модель на тестовом датасете
nb_classes <- predict(nb_fit, newdata = testing)
# оцениваем модель. Presision здесть обозначается как Pos Pred Value, а recall как Sensitivity
sink("NB.txt")  # сохраняем матрицу в txt-файл
time.taken
confusionMatrix(data = nb_classes, testing$myclass)
sink()

# SVM
start.time <- Sys.time()
svm_fit <- train(myclass ~ .,
                data = training, 
                method = "svmRadial",
                trControl = ctrl)
end.time <- Sys.time()
time.taken <- end.time - start.time
saveRDS(svm_fit, "./models-round/svm.RDS")
svm_classes <- predict(svm_fit, newdata = testing)
sink("svm.txt")
time.taken
confusionMatrix(data = svm_classes, testing$myclass)
sink()


# KNN
start.time <- Sys.time()
knn_fit <- train(myclass ~ .,
                 data = training, 
                 method = "kknn",
                 trControl = ctrl)
end.time <- Sys.time()
time.taken <- end.time - start.time
saveRDS(knn_fit, "./models-round/knn.RDS")
knn_classes <- predict(knn_fit, newdata = testing)
sink("knn.txt")
time.taken
confusionMatrix(data = knn_classes, testing$myclass)
sink()

# Loglit
#5 * ((ncol(testset) -1) + 1) + 5 + 1



start.time <- Sys.time()
logit_fit <- train(myclass ~ .,
                 data = training, 
                 method="multinom", MaxNWts = 21236, trControl=ctrl)
end.time <- Sys.time()
time.taken <- end.time - start.time
saveRDS(logit_fit, "./models-round/logit.RDS")