Machine Learning Basics in R

MACHINE LEARNING WORKFLOW:

Raw Data
    ↓
Data Preprocessing
  • Cleaning
  • Feature engineering
  • Scaling/normalization
  • Encoding categoricals
    ↓
Train/Test Split
  • 70-80% training
  • 20-30% testing
    ↓
Model Training
  • Choose algorithm
  • Set hyperparameters
  • Cross-validation
    ↓
Model Evaluation
  • Accuracy, RMSE, F1
  • ROC curve, AUC
    ↓
Prediction
  • Apply to new data

library(caret)
library(dplyr)

set.seed(42)
# Generate classification dataset: predict employee churn
n <- 500
employees <- data.frame(
  age         = sample(22:60, n, replace=TRUE),
  salary      = round(runif(n, 35000, 120000), -3),
  tenure      = sample(1:20, n, replace=TRUE),
  satisfaction= round(runif(n, 1, 10), 1),
  distance    = sample(1:50, n, replace=TRUE),
  dept        = sample(c("IT","HR","Finance","Sales"), n, replace=TRUE)
)
# Churn more likely: low satisfaction, long commute, low salary
employees$churn <- factor(
  rbinom(n, 1, prob = plogis(-3 + (-employees$satisfaction*0.3) +
                               (employees$distance*0.05) + (-employees$salary/50000))),
  labels = c("No", "Yes")
)

cat("Churn Distribution:\n")
print(prop.table(table(employees$churn)) * 100)

# One-hot encode categorical variables
employees_encoded <- dummyVars(~ ., data=employees %>% select(-churn),
                                fullRank=TRUE) %>%
  predict(employees %>% select(-churn)) %>%
  as.data.frame()
employees_encoded$churn <- employees$churn

# Train/Test split (80/20)
set.seed(42)
train_idx <- createDataPartition(employees_encoded$churn, p=0.8, list=FALSE)
train <- employees_encoded[train_idx, ]
test  <- employees_encoded[-train_idx, ]
cat(sprintf("Train: %d | Test: %d\n", nrow(train), nrow(test)))

# Feature scaling (for distance-based algorithms)
preProc <- preProcess(train %>% select(-churn), method=c("center","scale"))
train_scaled <- predict(preProc, train)
test_scaled  <- predict(preProc, test)

# Cross-validation setup (10-fold CV)
ctrl <- trainControl(
  method          = "cv",           # k-fold cross-validation
  number          = 10,             # 10 folds
  classProbs      = TRUE,           # Compute class probabilities
  summaryFunction = twoClassSummary,
  savePredictions = TRUE
)

# Train multiple models
set.seed(42)
# Logistic Regression
model_lr <- train(churn ~ ., data=train_scaled, method="glm",
                   family="binomial", trControl=ctrl, metric="ROC")
# k-Nearest Neighbors
model_knn <- train(churn ~ ., data=train_scaled, method="knn",
                    tuneLength=10, trControl=ctrl, metric="ROC")
# Random Forest
model_rf <- train(churn ~ ., data=train_scaled, method="rf",
                   tuneLength=5, trControl=ctrl, metric="ROC")

# Compare models
results <- resamples(list(LR=model_lr, KNN=model_knn, RF=model_rf))
summary(results)
dotplot(results, main="Model Comparison — Cross-Validated ROC")

# Predict on test set
pred_lr  <- predict(model_lr,  newdata=test_scaled)
pred_rf  <- predict(model_rf,  newdata=test_scaled)

# Confusion Matrix
cm_rf <- confusionMatrix(pred_rf, test_scaled$churn, positive="Yes")
print(cm_rf)

cat(sprintf("\nRandom Forest Performance:\n"))
cat(sprintf("  Accuracy:  %.3f\n", cm_rf$overall["Accuracy"]))
cat(sprintf("  Precision: %.3f\n", cm_rf$byClass["Precision"]))
cat(sprintf("  Recall:    %.3f\n", cm_rf$byClass["Recall"]))
cat(sprintf("  F1 Score:  %.3f\n", cm_rf$byClass["F1"]))
cat(sprintf("  Kappa:     %.3f\n", cm_rf$overall["Kappa"]))

# ROC Curve
prob_rf <- predict(model_rf, newdata=test_scaled, type="prob")[, "Yes"]
library(pROC)
roc_obj <- roc(test_scaled$churn, prob_rf, levels=c("No","Yes"))
cat(sprintf("  AUC:       %.3f\n", auc(roc_obj)))
plot(roc_obj, main="ROC Curve — Random Forest", col="#1565C0", lwd=2)

# Feature Importance
imp <- varImp(model_rf)
plot(imp, main="Feature Importance — Random Forest")