DETECTION OF FAKE AD CLICKS USING DATA-DRIVEN MACHINE LEARNING MODELS

Abstract

The proliferation of pay-per-click (PPC) digital advertising has given rise to a critical threat: fake ad click fraud, where bots or malicious actors generate illegitimate clicks causing billions of dollars in annual losses. This paper presents a hybrid machine learning system combining a supervised XGBoost classifier and an unsupervised Isolation Forest anomaly detector to identify fraudulent clicks in real time. Behavioral feature engineering extracts IP-level click counts, hourly burst patterns, and app-channel interaction frequencies from raw click-log data. The XGBoost model achieves 98.2% accuracy and an AUC-ROC of 0.974 with a scale_pos_weight of 438.56 to handle severe class imbalance. A weighted risk score (70% XGBoost + 30% Isolation Forest) drives a three-tier fraud classification: Genuine, Suspicious, and Fraud. A Streamlit dashboard provides real-time scoring, burst detection, feature importance visualization, and session export functionality