2nd Place · 25 teams
Song-Skip Prediction — MLDS Datathon
2026
Stacked ensemble hitting 79.6% accuracy on real-world streaming data.
The challenge
Given real-world music streaming data, predict whether a listener will skip a song. The signal is noisy and behavioral — the interesting work is squeezing generalizable structure out of messy human data under competition time pressure.
The approach
Instead of tuning a single model to death, my team built a stacked ensemble: LightGBM, XGBoost, and CatBoost gradient-boosted trees plus a Ridge regression baseline, with a meta-learner trained on their out-of-fold predictions. The meta-learner learns where each base model is trustworthy and weights them accordingly — the ensemble outperformed every individual model.
Result
79.6% accuracy, 2nd place out of 25 teams.
more projects