Abstract
Fraudsters in the real world frequently add more legitimate links while concealing their direct ones with other fraudsters, which leads to heterophily in fraud graphs, which is a problem that most GNN-based techniques are not built to solve. Several works have been proposed to tackle the issue in the spatial and spectral domains, but they are very limited due to a lack of understanding of spectral energy distribution in heterophily. In this paper, we analyze the spectral distribution with different heterophily degrees and observe that the heterophily of fraud nodes leads to the spectral energy moving from low-frequency to high-frequency. Further, we propose to split graphs using heterophilic and homophilic edges to capture signals in different frequency bands more effectively. We propose a novel spectral graph neural network, SplitGNN, to better capture signals for fraud detection against heterophily. SplitGNN uses an edge classifier to split the origin graph and adopts several band-pass spectral filters to learn representations. Extensive experiments on real-world datasets demonstrate the effectiveness of our proposed method. The code and data are available at https://github.com/SplitGNN/SplitGNN.
| Original language | English |
|---|---|
| Title of host publication | The 32nd ACM International Conference on Information and Knowledge Management |
| Publisher | ACM Digital Library |
| Publication status | Published - 2023 |