Data-Analytics/Lab 5/lab5.r

install.packages(
    c("e1071", "caret", "randomForest", "ggplot2", "pROC"),
    repos = c("https://cloud.r-project.org/"),
    dependencies = TRUE
)

suppressPackageStartupMessages({
    library(e1071) # for svm/tune.svm
    library(caret) # for metrics
    library(randomForest) # alternative classifier
    library(ggplot2)
})

set.seed(42)

read_wine <- function() {
    df <- read.csv("wine.data", header = FALSE)
    colnames(df) <- c(
        "Class",
        "Alcohol", "Malic.acid", "Ash", "Alcalinity.of.ash", "Magnesium",
        "Total.phenols", "Flavanoids", "Nonflavanoid.phenols", "Proanthocyanins",
        "Color.intensity", "Hue", "OD280.OD315", "Proline"
    )
    df$Class <- factor(df$Class)
    df
}

df <- read_wine()

# split into train/test
idx <- createDataPartition(df$Class, p = 0.8, list = FALSE)
train <- df[idx, ]
test <- df[-idx, ]

# choose a subset of features based on ANOVA F-test
# I picked this sbuset before the runs:
# alcohol, flavanoids, color intensity, od280/od315, proline, total phenols
features <- c("Alcohol", "Flavanoids", "Color.intensity", "OD280.OD315", "Proline", "Total.phenols")
x_train <- train[, features]
y_train <- train$Class
x_test <- test[, features]
y_test <- test$Class

# scale features
pp <- preProcess(x_train, method = c("center", "scale"))
x_train_s <- predict(pp, x_train)
x_test_s <- predict(pp, x_test)

# linear kernel svm with hyperparameter tuning (C)
set.seed(42)
lin_grid <- data.frame(cost = c(0.1, 1, 10, 100))
tune_lin <- tune.svm(
    x = x_train_s, y = y_train,
    kernel = "linear",
    cost = lin_grid$cost,
    tunecontrol = tune.control(cross = 5)
)
lin_best <- tune_lin$best.model

# rbf kernel svm with tuning (C, gamma)
set.seed(42)
rbf_grid_cost <- c(0.1, 1, 10, 100, 1000)
rbf_grid_gamma <- c(0.001, 0.01, 0.1, 1)
tune_rbf <- tune.svm(
    x = x_train_s, y = y_train,
    kernel = "radial",
    cost = rbf_grid_cost,
    gamma = rbf_grid_gamma,
    tunecontrol = tune.control(cross = 5)
)
rbf_best <- tune_rbf$best.model

# alt classifier: random forest (same features)
set.seed(42)
rf_fit <- randomForest(x = x_train, y = y_train, ntree = 500, mtry = 2, importance = TRUE)

# evaluation helper
eval_model <- function(model, x_test_s, y_test, name) {
    pred <- predict(model, x_test_s)
    cm <- confusionMatrix(pred, y_test)
    pr <- data.frame(
        model = name,
        accuracy = cm$overall["Accuracy"],
        precision_macro = mean(cm$byClass[, "Precision"], na.rm = TRUE),
        recall_macro = mean(cm$byClass[, "Recall"], na.rm = TRUE),
        f1_macro = mean(cm$byClass[, "F1"], na.rm = TRUE)
    )
    list(cm = cm, pr = pr)
}

# eval svm models (use scaled features)
lin_eval <- eval_model(lin_best, x_test_s, y_test, "svm_linear")
rbf_eval <- eval_model(rbf_best, x_test_s, y_test, "svm_rbf")

# evaluate random forest (no scaling)
rf_pred <- predict(rf_fit, x_test)
rf_cm <- confusionMatrix(rf_pred, y_test)

rf_pr <- data.frame(
    model = "random_forest",
    accuracy = rf_cm$overall["Accuracy"],
    precision_macro = mean(rf_cm$byClass[, "Precision"], na.rm = TRUE),
    recall_macro = mean(rf_cm$byClass[, "Recall"], na.rm = TRUE),
    f1_macro = mean(rf_cm$byClass[, "F1"], na.rm = TRUE)
)

perf <- rbind(lin_eval$pr, rbf_eval$pr, rf_pr)

# print
cat("best params (linear svm): C =", lin_best$cost, "\n")
cat("best params (rbf svm): C =", rbf_best$cost, " gamma =", rbf_best$gamma, "\n\n")
print(perf)

# macro-f1 comparison
ggplot(perf, aes(x = model, y = f1_macro)) +
    geom_col() +
    labs(title = "macro-F1 by model (wine test set)")

# save outputs
write.table(perf, file = "lab5_performance_table.txt", sep = "\t", row.names = FALSE, quote = FALSE)
sink("lab5_confusion_matrices.txt")
cat("=== svm linear ===\n")
print(lin_eval$cm)
cat("\n=== svm rbf ===\n")
print(rbf_eval$cm)
cat("\n=== random forest ===\n")
print(rf_cm)
sink()