TalkingData AdTracking Fraud Detection

Kaggle project to identify fraudulent clicks for mobile advertising. I ranked in the top 10% out of 3,967 teams on the public leaderboard.

The raw data contained more than 180 million click records over three days. I engineered dozens of features from click counts and time intervals across combinations of IP address, OS, app, device, and ad publisher. LightGBM gradient boosting models achieved AUC-ROC above 0.981.

Project repoKaggle competition