Model Comparison — Original Data
What you'll see here
Train 5 models on the original (imbalanced) training data and compare Accuracy / Precision / Recall / F1 / AUC on the held-out test set.
Training controls
5 models · Test = 30% stratified
Larger training subsamples give better metrics but take longer. Trees: 25 for ensembles (single tree for DT), maxDepth 8 (1 for AdaBoost stumps, 3 for GBM), lr=0.1 for GBM.