The Devil is in the Conflict: Disentangled Information Graph Neural Networks for Fraud Detection

TL;DR

This paper introduces DIGNN, a graph neural network that disentangles topology and attribute information with attention and mutual information constraints, effectively addressing the homophily assumption issue in fraud detection.

Contribution

The paper proposes a novel disentangled GNN model with attention and mutual information constraints to improve fraud detection under heterophily conditions.

Findings

DIGNN significantly outperforms state-of-the-art methods on real-world datasets.

Disentangling topology and attribute improves detection accuracy.

Mutual information constraints enhance the model's robustness.

Abstract

Graph-based fraud detection has heretofore received considerable attention. Owning to the great success of Graph Neural Networks (GNNs), many approaches adopting GNNs for fraud detection has been gaining momentum. However, most existing methods are based on the strong inductive bias of homophily, which indicates that the context neighbors tend to have same labels or similar features. In real scenarios, fraudsters often engage in camouflage behaviors in order to avoid detection system. Therefore, the homophilic assumption no longer holds, which is known as the inconsistency problem. In this paper, we argue that the performance degradation is mainly attributed to the inconsistency between topology and attribute. To address this problem, we propose to disentangle the fraud network into two views, each corresponding to topology and attribute respectively. Then we propose a simple and effective method that uses the attention mechanism to adaptively fuse two views which captures data-specific preference. In addition, we further improve it by introducing mutual information constraints for topology and attribute. To this end, we propose a Disentangled Information Graph Neural Network (DIGNN) model, which utilizes variational bounds to find an approximate solution to our proposed optimization objective function. Extensive experiments demonstrate that our model can significantly outperform stateof-the-art baselines on real-world fraud detection datasets.