An Empirical Analysis of Click Temporal Features for Automated Ad Fraud Detection in Mobile In-App Browser Environments

Abstract

Mobile in-app browsers (IABs) embedded inside social, messaging, and content applications now serve a substantial share of digital advertising impressions, and they have become an attractive surface for click manipulation generated by automated scripts and click farms. Building reliable detectors that respect user privacy requires a clear understanding of which features actually carry discriminative signal. This paper presents an empirical analysis of click temporal features (those derived solely from event timestamps) for distinguishing invalid from legitimate ad clicks in IAB-style traffic. Using the public TalkingData AdTracking Anomaly Detection dataset of approximately 184 million labelled clicks, complemented by the Avazu mobile-advertising corpus and the HuMIdb behavioural dataset, we extract four families of temporal features: inter-arrival times, burstiness statistics, time-of-day aggregation, and short-window periodicity. We characterise each family using non-parametric distribution comparisons and quantify its standalone discriminative power inside two well-established gradient-boosting classifiers. The analysis shows that simple inter-arrival statistics already account for the majority of attainable AUC, that adding burst and aggregation features yields measurable but small gains, and that detectors degrade systematically under behaviour mimicry. The findings provide IAB operators with a reproducible baseline for prioritising lightweight, privacy-friendly temporal signals.