How a Rotten Apple Can Spoil the Barrel: Corporate Fraud Detection via Dynamic Business Networks

Corporate fraud detection is attracting renewed research interest in the era of internet connectivity due to its increasing complexity, concealment, and long-term duration. Previous fraud detection methods have heavily relied on financial statements and demonstrated limited detection power. While network analysis has emerged as a promising technique, its full potential in corporate fraud detection remains largely untapped, thus missing out on the potential improvement of fraud detection power. Recognizing this significant research gap, we devised a novel framework, NetDetect, to jointly learn structural information and temporal changes in dynamic heterogeneous networks for effective and interpretable corporate fraud detection. NetDetect consists of two new modules—meta-path-based static network feature extraction (SNFE) and temporal network feature extraction (TemFE). We rigorously evaluated each module and the entire framework against state-of-the-art baselines using a unique dataset of Chinese listed companies with a 561K node network. Experiment results and a case study demonstrate that our framework detects corporate fraud more accurately compared to baselines and offers good early-warning capability, as well as results interpretability. Our research results have important implications for various stakeholders (e.g., auditors, investors, and regulators) seeking to understand and contain corporate fraud risk. We also discuss contributions to the IS knowledge base with two design principles to guide the development of similar IT artifacts.