CSIP: Enhanced Link Prediction with Context of Social Influence Propagation

Abstract

Data mining in social networks brings an indispensable role for the construction of smart cities from the perspective of social development. Link prediction is an important task of data mining, especially in the knowledge graph, which is also called knowledge graph completion. Link prediction aims to find missing links or predict potential links according to the current social network. The most existing link prediction methods focus on static information in social networks, such as topology and node attributes, which are partly provided by users. When users are unwilling to provide or intentionally hide these static features, traditional link prediction methods cannot achieve ideal performance. The dynamic information of social influence propagation in social networks can avoid the user's subjective impact and better reflect the relationship between users. In addition, users show different degrees of interest and authority on various topics in the real world, leading to different influence propagation patterns. Therefore, we use context of social influence to optimize the topic-aware influence propagation model to improve the performance of link prediction. In this paper, we propose a new multi-output graph neural network framework to capture influence propagation in social networks and model the influence of users in different roles. In this way, the underlying information of influence between users can be used to construct new features to improve the performance of link prediction. Our experiments conduct the method on multiple benchmark datasets. The experimental results show that the modeling of context is effective, and our model outperforms the compared state-of-the-art link prediction methods.