Modeling Connection Strength in Graph Neural Networks for Social Influence Prediction

Social media applications such as Twitter and Weibo have attracted tremendous interests in understanding and predicting online social influence. Among other factors, relationship closeness between users is one of the most important ones for social influence prediction. Recent studies have verified the effectiveness of employing graph neural networks(GNN) for social influence prediction. However, most of them are failed to or less effective to make use of relationship closeness explicitly and precisely (with a few exceptions in which relationship closeness is measured as connection strength). Instead, they tend to use simple binary information such as whether users are friends/followers or not. In this paper, we develop a new model, CSGNN, to exploit connection strength feature in GAT and GCN. We estimate connection strength by a combination of static interaction frequency and time duration of follow actions. We also propose a new attention mechanism to incorporate connection strength based on GAT, and a new method of convolution and aggregation based on GCN. Extensive experiments are conducted on multiple real-world social network datasets, including Digg and Weibo. The results demonstrate that CSGNN outperforms several state-of-the-art baselines, and the ablation study confirms that connection strength contributes to the performance improvement significantly.