stys / kaggle-talkingdata-adtracking-fraud-detection Goto Github PK

Kaggle TalkingData Ad Tracking Fraud Detection Challenge

Scripts for competition https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection

Final ranking #55 with private leaderboard score 0.9824250

• Aggregate factors from Baris' kernel computed on all data with test_supplement
• Prev/next click times (1-, 2-, 3-step) computed on all data with test_supplement
• LIBFFM out-of-fold using 5 folds
• LGBM on 50% data (using all is_attributed=1 and 50% of is_attributed=0)
• Average of a few LGBM models trained on different subsets of data and factors
• Training with 96GB RAM
• Store data column-wise in binary formats for fast loading
• HOCON configurations are awesome

EDA preprocessing/eda.ipyndb - distributions of frequencies in train.csv

Merge test datasets preprocessing/merge_test_sets.ipynb - script for merging test datasets test.csv test_supplement.csv

Baris kernel https://www.kaggle.com/bk0000/non-blending-lightgbm-model-lb-0-977?scriptVersionId=3224614 https://www.kaggle.com/aharless/kaggle-runnable-version-of-baris-kanber-s-lightgbm

Next/prev clicks without hashing trick https://www.kaggle.com/asydorchuk/nextclick-calculation-without-hashing-trick

Submissions

Factors strength (lgbm_19)