Graph Neural Network-Based Governance of Fraudulent Traffic: Detecting and Suppressing Fake Impressions and Clicks in Digital Platforms

Abstract

Digital advertising platforms rely heavily on impression and click metrics to quantify user engagement, but fraudulent traffic-including fake impressions generated by bot farms and click farming-causes over $68 billion in annual losses for global advertisers. Traditional anti-fraud methods, such as rule-based engines and isolated machine learning (ML) models, fail to capture the complex relational patterns among users, advertisements, and traffic sources, leading to high false positive rates and poor adaptability to evolving fraud tactics. To address these gaps, this study proposes a Dynamic Graph Neural Network-based Fraud Traffic Governance Framework (DGNN-FTG) that integrates real-time detection and adaptive suppression in a unified pipeline. First, we construct a User-Advertisement-Medium (UAM) dynamic interaction graph that encodes spatio-temporal behavioural features and relational dependencies, overcoming the limitations of static, single-node modelling. Second, we design a Spatio-Temporal Dynamic GNN (ST-DGNN) for fraud detection, which incorporates temporal attention to track behaviour evolution and spatial attention to identify anomalous relational clusters. Third, we develop an adaptive suppression module that adjusts traffic filtering thresholds based on detection confidence, combined with a false-positive compensation mechanism to minimize disruption to legitimate users. Validated on a real-world dataset from a leading e-commerce advertising platform, DGNN-FTG achieves a fraud detection F1-score of 92.8%, outperforming traditional XGBoost and static GNN models by 18.3% and 11.5%, respectively. The suppression module reduces fraudulent traffic by 89.2% while maintaining a false positive rate of only 2.1%, balancing anti-fraud efficacy and user experience. This framework provides a scalable, real-time solution for digital platforms to combat fraudulent traffic and safeguard the integrity of the advertising ecosystem. Notably, DGNN-FTG exhibits strong generalization across platform types, achieving an F1-score of over 90% when deployed on both e-commerce and social media advertising systems, with minimal retraining required for cross-platform adaptation. For small and medium-sized advertisers, the framework's low false positive rate translates to a 22.3% reduction in wasted ad expenditure, directly improving their return on investment (ROI) in digital marketing campaigns. Additionally, the dynamic graph update mechanism enables DGNN-FTG to detect emerging fraud tactics-such as AI-generated fake user behaviour-within 48 hours of their appearance, far faster than the 7-10 day response time of traditional anti-fraud systems.