Adfraud System: Real-Time Ad Click Fraud Detection Using Stacking Ensemble, Deep Learning, and an Agentic AI Chatbot

Abstract

Ad click fraud drains billions of advertiser budgets annually through bots and click farms that generate fake clicks with zero genuine engagement. This paper presents Adfraud system, a production-ready fraud detection system combining a novel 18-signal real-time feature engineering engine with nine ML/DL algorithms and an agentic AI chatbot. Operating on the public TalkingData AdTracking benchmark (100,000 records; 0.227% positive class), the system engineers fraud signals from raw click telemetry — click burst velocity, device–OS consistency, impossible geolocation, subnet botnet flags, and user-agent entropy — feeding a Stacking Classifier (LR+RF+XGBoost+LightGBM → meta-LR) achieving 97.4% accuracy, 96.8% F1, and AUC 0.98 — statistically significantly outperforming all eight baselines (Friedman χ²=47.3, p<0.0001). SHAP attribution identifies impossible geolocation and device–OS mismatch as the strongest discriminators. The deployed Flask platform exposes 20 REST endpoints, SSE live monitoring, batch processing, model drift detection, multi-website API-key tracking, and an agentic AI chatbot with six specialised fraud-analysis tools. The system is fully containerised via Docker.