Addressing Noise and Stochasticity in Fraud Detection for Service Networks

TL;DR

This work tackles fraud detection in complex service networks by addressing noise in information propagation and the lack of frequency-specific discrimination in graph signals. It introduces SGNN-IB, a spectral GNN that splits the graph into homophilic and heterophilic subgraphs using an edge classifier, applies Beta wavelet band-pass filters to capture low- and high-frequency signals, and fuses them with prototype learning. An information-bottleneck (IB) based denoising module and joint training objective improve representation quality and robustness against noise, while a heterophily-aware edge classifier preserves discriminative structure. Extensive experiments on three public datasets show SGNN-IB consistently outperforms state-of-the-art methods, validating the effectiveness of frequency-aware fusion and IB-driven denoising for fraud detection in multi-relational service networks.

Abstract

Fraud detection is crucial in social service networks to maintain user trust and improve service network security. Existing spectral graph-based methods address this challenge by leveraging different graph filters to capture signals with different frequencies in service networks. However, most graph filter-based methods struggle with deriving clean and discriminative graph signals. On the one hand, they overlook the noise in the information propagation process, resulting in degradation of filtering ability. On the other hand, they fail to discriminate the frequency-specific characteristics of graph signals, leading to distortion of signals fusion. To address these issues, we develop a novel spectral graph network based on information bottleneck theory (SGNN-IB) for fraud detection in service networks. SGNN-IB splits the original graph into homophilic and heterophilic subgraphs to better capture the signals at different frequencies. For the first limitation, SGNN-IB applies information bottleneck theory to extract key characteristics of encoded representations. For the second limitation, SGNN-IB introduces prototype learning to implement signal fusion, preserving the frequency-specific characteristics of signals. Extensive experiments on three real-world datasets demonstrate that SGNN-IB outperforms state-of-the-art fraud detection methods.

Figures (7)

• Figure 1: The framework of SGNN-IB. First, SGNN-IB leverages an edge classifier to perceive heterophilic subgraphs. Then, SGNN-IB utilizes multi-scale graph filters to obtain the high- and low-frequency signals in the graph. Subsequently, SGNN-IB integrate the signals from different frequency based on prototype learning. Finally, SGNN-IB is trained by the joint loss function, integrated with IB-loss.
• Figure 2: The architecture of IB loss. To solve the noise issue, the model leverages classical IB theory, maximizing the mutual information between the latent features and the ground truths and minimizing the mutual information between the latent features and the original features. Here, latent features denote the high-pass and low-pass signals and ground truths represent the band-pass signals. To solve the stochasticity issue, the model introduces the mutual information between high-pass and low-pass signals.
• Figure 3: Sensitivity experimental results on YelpChi dataset: (a) Sensitivity results for parameter $\lambda$; (b) Sensitivity results for parameter $\eta$; (c) Sensitivity results for parameter $\mu$.
• Figure 4: Sensitivity experimental results on the Amazon dataset: (a) Sensitivity results for parameter $\lambda$; (b) Sensitivity results for parameter $\eta$; (c) Sensitivity results for parameter $\mu$.
• Figure 5: Sensitivity experimental results on FDCompCN dataset: (a) Sensitivity results for parameter $\lambda$; (b) Sensitivity results for parameter $\eta$; (c) Sensitivity results for parameter $\mu$.