SMART ADVERTISEMENT CLICK FRAUD DETECTION SYSTEM

Abstract

Predicting and detecting mobile advertisement click fraud is a critical challenge for digital marketing infrastructure. This paper presents a comprehensive machine learning framework that benchmarks fourteen supervised classifiers on the TalkingData Ad Tracking Fraud Detection dataset. A balanced corpus of 913,692 samples is constructed through majority-class under-sampling over seven click-level features: IP address, application ID, device type, operating system, channel, hour, and minute. Eight classical machine learning algorithms, three deep learning architectures (ANN, CNN, RNN), and three proposed models — a Stacking Classifier, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) — are evaluated under identical experimental conditions. The proposed Stacking Classifier, combining XGBoost, Random Forest, Decision Tree, Gradient Boosting, and LightGBM with a Logistic Regression meta-learner, achieves 92% accuracy and 89% fraudulent-class recall — the highest of all evaluated models. GRU achieves 91% accuracy with 0.91 F1-score at lower computational cost than LSTM, making it the recommended deployment architecture. Results confirm that heterogeneous ensembles and recurrent networks substantially outperform linear classifiers for this non-linear, imbalanced classification task.