Random Forests and Overfitting: Understanding Behavior and Performance
Random forests learn local probability peaks, yielding near-perfect training c-statistics, but modest discrimination loss on test data. Avoiding deeply grown trees may lead to better performance.