利用异配及同配关系的社交机器人检测方法

Due to can not differentiate the heterophilic and homophilic relations between users, the existing graph-based methods can lead to noise in user representations of heterophilic edges, diminishing the differences between social robots and human users, resulting in a decrease in detection accuracy. Social robots continuously evolve to mimic human user behaviors such as the number of followers and the number of tweets posted, leading to an increased proportion of sophisticated social robots, reducing the detection recall rate. In this paper, a social robot detection method was proposed based on both heterophilic and homophilic relations, employing memory network for constructing prototypes of heterophilic and homophilic relations to identify user relation types, reducing the interference of heterophilic edges on user representations and enhancing the feature distinctiveness between different types of users. A regulatory factor was introduced into loss function to enhance the loss contribution of hard-to-classify users during the parameter optimization process, improving the ability of model to recognize sophisticated social robots. Experimental results show that the proposed method can outperform current advanced approaches in distinguishing heterophilic and homophilic relations between users, reducing the weight of heterophilic edges, enhancing the class separability of user representations, and improving the detection accuracy effectively even in cases of low homophily scores.