Conspiracy Spoofing Detection via Structure-Augmented Generative Graph Model

Abstract

Detecting spoofing in financial trading is a critical data mining task. Traditional machine learning models often focus on individual node features, failing to capture the contextual relationships among interconnected nodes. Graph-based methodologies have enhanced this by effectively integrating relational data. Recent advancements in fraud detection demonstrate substantial performance gains by incorporating structure information into detection models. However, spoofing transactions often exhibit a distribution shift compared to historical transactions, rendering historical data less effective. Instead, certain trading patterns, such as motif structures, consistently manifest in transaction graphs regardless of distribution shift, providing a robust alternative for analysis. Motif structures, particularly node motifs, are essential for capturing higher-order interactions and structural patterns within transaction graphs. This paper introduces the Structure-Augmented Generative Graph Model (SAG2M) to address the challenge of detecting conspiracy spoofing through a substructure frequency-augmented detection method. Specifically, our approach extracts the frequency of subgraph patterns among neighboring nodes, leveraging an enumeration algorithm to efficiently identify node orbit data. The extracted motif frequencies are then encoded into a structure-augmented generative framework, enabling detailed structural representations of each transaction (node). Subsequently, a temporal and heterogeneous graph generation and aggregation scheme is applied to collect neighborhood node information, uncovering conspiracy spoofing patterns effectively. Our experiments on datasets such as Amazon, Yelpchi, and T-Finance demonstrate that SAG2M outperforms existing models in detection accuracy. A case study focusing on conspiracy spoofing detection further highlights the model’s superior effectiveness in identifying such complex fraudulent behaviors.