Abstract:

As social bot technology advances, these bots increasingly interact with human users, making their detection a more challenging problem. Existing detection methods primarily rely on the homophily assumption, often overlooking the connections between different classes of users, particularly the impact of heterophily. This oversight impairs their detection performance. To address this issue, this paper presented an innovative social bot detection framework that emphasizes social heterophily. It leveraged user connections within social networks and extensively explored various types of social information to mitigate the effects of heterophily and achieved more accurate detection. This paper examined user relationships from both homophily and heterophily perspectives. It constructed the social network as a graph and employed a message-passing mechanism to aggregate information from both homophilic and heterophilic edges, allowing for the extraction of frequency-based node features. Furthermore, it aggregated features from various nodes within the graph to generate social context features. These features are then blended and utilized for the detection task. The experimental results validate the method’s superiority over comparative approaches on publicly available datasets, confirming its effectiveness.

Key words: social bot detection, homophily and heterophily, graph neural network