Towards self-interpretable review spammer group detection

Due to data imbalance, existing spammer group detection methods often yield suboptimal performance. Moreover, many of these approaches operate as black boxes, offering little to no interpretability for their detection results. To address these issues, we propose a self-interpretable approach for detecting review spammer groups. Our method begins by constructing a user-product bipartite graph with edge attributes. We integrate user review information with a novel fitness function to develop an adaptive genetic algorithm that effectively identifies high-quality candidate groups. Next, we introduce a hybrid graph neural network enhanced with active learning to generate vector representations of these candidate groups. We then design and construct a prototype layer and a group classification layer to detect spammer groups accurately. To provide interpretability, we incorporate prototype learning to create an interpretation mechanism that explains detection outcomes. Experimental results demonstrate that our method achieves substantial improvements in Precision@k and Recall@k at the top-1000 ranking, outperforming state-of-the-art solutions on the Amazon, YelpChi, YelpNYC, and YelpZip datasets by [11.53 %, 97.36 %, 51.37 %, 32.06 %] and [12.65 %, 54.65 %, 64.19 %, 47.37 %], respectively. Additionally, the Fidelity of our interpretability results under varying Sparsity levels is approximately 4 %, 9 %, 8 %, and 8 % higher than those of existing methods on the same datasets.