Dual-Channel Relation Mixup Strategy for Social Bot Detection Based on Graph Neural Network

Abstract In social bot detection, accurate labeling needs to be done manually, resulting in large-scale labels are difficult to obtain. However, the limited labels mean that there is less data available for the training set, which may result in over-fitting. Therefore, we proposed a novel data augmentation strategy named Dual-Channel Relation Mixup (DCRM) to address potential over-fitting issues caused by a small training set. In our strategy, we firstly construct two channels that input different account relations, then combine the resulting aggregated representations from both channels before passing them on to the next layer of the “message passing” mechanism. More critically, our DCRM is a general skill that can be equipped with many other backbone models (e.g., GCN, GraphSAGE, GAT, and RGCN) to enhance performance. Extensive experiments on several benchmarks demonstrate that our proposed DCRM strategy is an effective way to enhance the performance of bot detection based on graph neural networks. Furthermore, it also exhibits better generalization ability and stability, especially when working with smaller datasets.