PEGNN: Peripheral-Enhanced graph neural network for social bot detection

Abstract

The development of social networks plays a stronger role in the increased importance of social bot detection. One of the most common existing graph-based detection methods is graph neural networks (GNNs). However, researchers mainly focus on the improvement of GNN architecture while neglecting the in-depth analysis of social network structure. In this paper, after in-depth analysis of the social graph structure, it demonstrate that there exists a graph stratification phenomenon, i.e., the social graph is divided into central and peripheral layers according to its topology. Based on this phenomenon, the idea that the social robot detection task should focus on the central node is proposed, and the existing detection task is adapted to a central node classification task. The task is then further investigated, and the Peripheral-Enhanced Graph Neural Network (PEGNN) framework is proposed to tackle the problem that existing frameworks cannot effectively utilize the information of peripheral networks. PEGNN effectively utilizes the information of peripheral networks via the synergistic effect of three losses. Eventually, the graph stratification phenomenon is verified on two datasets, and the original task is adjusted for central node classification to verify the enhancement effect of PEGNN. The experimental results exhibit that PEGNN obviously improves the model performance. On Twibot-20, the average improvement in accuracy is 1.30%, and the average improvement in F1 score is 1.37%; on Twibot-22, the average improvement in accuracy is 2.46%, the average improvement in F1 score is 5.91%; apparently, other metrics also show obvious improvement.