LightGBM random search
Single-family LightGBM tuned by random search. Trains in minutes, gets within 0.002 AUC of the full stacked ensemble on gender, and is the right answer when you do not want to manage five separate gradient boosting libraries.
Scores
How it works
The number reported here is the best single LightGBM configuration out of the 300-trial LGB random search the GBM stack already runs, refit with the same 5-fold CV so the OOF predictions are honest. The search space covers learning rate, number of leaves, max depth, regularisation (L1 / L2 / min split gain), column and row subsampling, and the n_estimators budget.
The interesting observation is that for the mood head the random search actually loses to a hand-picked default configuration (8.85 MAE). 30 trials is not always enough to beat a well-chosen starting point on a hard target.
Configuration
- FamilyLightGBM only
- Objectivebinary (gender) · regression_l1 / MAE (age, mood)
- Search size300 random samples per target (from the GBM stack search)
- ReportingBest single configuration, refit under shared CV
- CV5-fold, seed 42