Single-model baselines
Five hand-tuned single configurations under the shared 5-fold CV. No random search, no stacking, no per-target tweaking - one default config per family that a reasonable practitioner would write on day one. The job of this page is to keep the rest of the leaderboard honest: anything that does not beat these by a clear margin is not earning its keep.
Per-model scores
| Model | Gender AUC | Age MAE | Mood MAE |
|---|---|---|---|
| Single LightGBM | 0.876 | 6.95 | 8.85 |
| Single XGBoost | 0.875 | 7.14 | 9.08 |
| HistGradientBoosting | 0.867 | 7.16 | 9.27 |
| MLP (256, 128, 64) | 0.823 | 8.41 | 10.84 |
| Linear (logistic / ridge) | 0.810 | 8.77 | 11.35 |
Best baseline per column highlighted. A simple LightGBM with sensible defaults wins every metric here. The blended baseline row on the leaderboard (logistic + HGB + LGB + XGB + MLP) lands at 0.874 / 7.10 / 9.10 because the linear and MLP members drag the trees down.
Configurations
- Logistic / ridgeStandardScaler + sklearn LogisticRegression (gender) or Ridge (regression), default alpha
- HistGradientBoostingmax_iter=600, max_depth=7, learning_rate=0.05, l2_regularization=0.5
- LightGBMn_estimators=1000, num_leaves=63, learning_rate=0.03, feature_fraction=0.8, bagging_fraction=0.8, bagging_freq=5
- XGBoostn_estimators=1000, max_depth=6, learning_rate=0.03, subsample=0.85, colsample_bytree=0.85
- MLPStandardScaler + hidden=(256,128,64), max_iter=400, alpha=1e-3, early stopping on 10% val
- Feature vectorSame 441 engineered features as the tree-based architectures (full breakdown)
- CV5-fold StratifiedKFold (gender), KFold (age, mood), seed 42