Machine Learning Workflows

Beyond classical statistics, R offers a full machine learning toolkit. This lesson covers unsupervised clustering, the Random Forest ensemble method, and how to properly evaluate any classification model.

1. K-Means Clustering (Unsupervised Learning)

K-means groups similar observations into k clusters without any labeled outcome variable.

set.seed(42)
df_scaled <- scale(df[, c("age", "salary")])  # standardize features first!

km_model <- kmeans(df_scaled, centers = 3, nstart = 25)
df$cluster <- km_model$cluster

table(df$cluster)
km_model$centers   # coordinates of each cluster center

2. Choosing the Number of Clusters — Elbow Method

wss <- sapply(1:10, function(k) {{
  kmeans(df_scaled, centers = k, nstart = 10)$tot.withinss
}})

plot(1:10, wss, type = "b", pch = 19,
     xlab = "Number of Clusters (k)", ylab = "Total Within-Cluster Sum of Squares")

3. Random Forest — Supervised Classification/Regression

Random Forest builds many decision trees and combines their votes — robust, accurate, and requires minimal tuning.

install.packages("randomForest")
library(randomForest)
library(caret)

set.seed(42)
train_index <- createDataPartition(df$churned, p = 0.8, list = FALSE)
train_data <- df[train_index, ]
test_data  <- df[-train_index, ]

rf_model <- randomForest(as.factor(churned) ~ tenure + monthly_charges + support_calls,
                          data = train_data, ntree = 500, importance = TRUE)

print(rf_model)
importance(rf_model)          # which variables matter most
varImpPlot(rf_model)          # visualize variable importance

4. Making Predictions with Random Forest

predictions <- predict(rf_model, newdata = test_data)

5. Model Evaluation — Confusion Matrix

library(caret)

conf_matrix <- confusionMatrix(predictions, as.factor(test_data$churned))
print(conf_matrix)

6. Model Evaluation — ROC and AUC

install.packages("pROC")
library(pROC)

prob_predictions <- predict(rf_model, newdata = test_data, type = "prob")[, 2]
roc_obj <- roc(test_data$churned, prob_predictions)

plot(roc_obj, main = "ROC Curve")
auc(roc_obj)   # Area Under the Curve — closer to 1 is better

7. Train/Test Split Best Practice

• Always split data before any preprocessing that "learns" from the data (like scaling parameters).
• A common split is 70–80% training, 20–30% testing.
• Use set.seed() so your split — and results — are reproducible.

8. Production-Ready Checklist

• ✅ Scale features before k-means — and choose k with the elbow method.
• ✅ Use Random Forest for robust classification — minimal tuning, strong baseline.
• ✅ Split data into train/test — before evaluating any supervised model.
• ✅ Evaluate with confusion matrix + ROC/AUC — not accuracy alone.