Kaggle TalkingData AdTracking Fraud Detection Challenge

48th Solution, competition link: https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection

1. features.ipynb: notebook version
2. train_xgb_lgb.py: script version, gives about 0.9824 on private LB
3. blending.ipynb: blending historical models, which boost me about 0.0002
4. FTRL.ipynb: haven't tried due to limited time

running this code on full training data needs 96GB RAM with 128G swap

Some Solutions as a Reference

1. 3th, NN model
2. 4th
3. 6th, strong single lightGBM
4. 9th
5. FFM trick

Train log

please see the dashboard