Detecting Camouflaged Social Bots Through Multi-level Aggregation and Information Encoding

Social bot detection has become a crucial task for maintaining order on social platforms. Many existing works use graph-based methods to enhance social bot detection. Although these graph-based methods have performed well, they ignore relation camouflage, where bot accounts interact with a large number of human accounts. In addition, the rich label, structure, and relation information contained in the original multi-relational graph is under-utilized. To overcome the above challenges, we propose a graph-based social bot detection framework named MAIE. Specifically, we first group the neighbors based on their labels and generalize this to higher-order neighbors with multiple relations to generate distinguishable neighborhood information. Then, we design three information encodings to retain the label, structure, and relation information of user nodes. Finally, using prior information from class-center, we introduce weight-guided loss to increase the discriminability between humans and bots. Experimental results demonstrate that MAIE achieves state-of-the-art performance on three mainstream social bot detection datasets.