Project · 2025
Depression Prediction (Kaggle)
Tabular-ML Kaggle entry predicting depression from survey features. Gradient-boosted trees, threshold tuning, and an evaluation harness that survived class imbalance.
Kaggle competition entry on depression prediction from survey features. The headline approach was gradient-boosted trees with threshold tuning and a small ensemble, but the interesting work was upstream of the model.
The dataset is severely imbalanced, which makes it very easy to get a deceptively good score by leaning into the majority class. Most of my time went into building an evaluation harness I could actually trust — stratified resampling, properly held-out folds, and a leaderboard of submission attempts so I could see which changes actually helped.