A hybrid feature-based heterogeneous graph transformer method for cheating official account mining

As cheating official accounts (COAs) in social media has posed a significant threat to society, how to detect COAs has attracted increasing attention worldwide. However, mining COAs is challenging because the COAs are always much fewer than normal accounts and will conceal themselves by mimicking normal accounts. Therefore, we propose a novel hybrid heterogeneous graph transformer (HHGT) method with both content and topology features to mine COAs more accurately, with two main contributions. First, we tackle the COA mining problem for the first time with a real-world dataset established from the WeChat platform, where the data analysis reveals that not only the node content feature but also the node topology feature can be useful for COA mining. Second, we propose the HHGT by learning both node content and topology features to exhibit superior generalization and representation capabilities for COA mining. We benchmark the HHGT and study its performance and feature effectiveness on the real WeChat dataset, providing helpful insights for COA mining. Experimental results show that the proposed HHGT performs better than all compared methods on unseen data, and can provide support for the analysis of the relationship among official accounts.