// the find
wepe/O2O-Coupon-Usage-Forecast
1st Place Solution for O2O Coupon Usage Forecast
A winning solution from a 2016 Alibaba Tianchi competition predicting whether users will redeem O2O coupons within 15 days of receiving them. It's a classic Kaggle-style ensemble of XGBoost, GBDT, and Random Forest on hand-engineered features. Useful as a reference for tabular ML feature engineering on user/merchant interaction data, not as production code you'd actually run.
The feature engineering is genuinely thorough — they break down user, merchant, coupon, and user-merchant interaction features systematically, which is a good template for anyone doing similar behavioral prediction work. The README honestly documents the data leakage they exploited, which is rare and useful for understanding competition-vs-reality gaps. The blending approach is clearly explained with specific weights (0.65 GBDT + 0.35 XGBoost) rather than vague gestures at 'model fusion'. Three separate contributors means you get three different implementation styles for the same problem, which is useful for comparison.
This is 2016 competition code frozen in time — scikit-learn and XGBoost APIs have changed enough that it won't run without fixes, and there are no requirements files or environment specs. The core leakage features that drove 10 AUC points of improvement are explicitly impossible to use in production, so the best-performing version of this solution is irreproducible in any real business context. The code across three contributors is not unified — you get three separate codebases with no shared data pipeline, which makes it hard to follow end to end. No data is included and the download link points to a Baidu pan URL that almost certainly no longer works.