vancenceho/vcp-combined-ensemble-stacking

Binary classifier for track-level YouTube virality (viral) from Spotify + Librosa audio features; stacking combines tuned RF, XGBoost, and LightGBM with a logistic meta-learner.

Stacking ensemble (combined features)

• Level 0: tuned Random Forest, XGBoost, and LightGBM, each producing P(viral) on the same 95-feature inputs (Spotify + Librosa audio; no direct YouTube engagement inputs).
• Level 1: logistic regression combines those three probabilities.

The uploaded stacking_ensemble_combined.joblib artifact is only the logistic meta-learner; production inference requires rf_combined.joblib, xgb_combined.joblib, and lgbm_combined.joblib plus feature_info_combined.joblib (and the same preprocessing as training). Primary reported test metrics for the full stack: AUC ≈ 0.79, accuracy ≈ 0.80 on the stratified holdout split.

Deployment:

1. Load feature_info_combined.joblib for column order
2. Rebuild X with the same exclusions
3. Apply the same mean imputation as training
4. Load rf_combined.joblib, xgb_combined.joblib, lgbm_combined.joblib, and stacking_ensemble_combined.joblib
5. Inference pipeline = three predict_proba[:,1] → stack → meta-learner.