Group link prediction in bipartite graphs with graph neural networks

Group link prediction is of both theoretical and practical significance since it can be used to analyze relationships between individuals and groups. However, obeying the homophily assumption, most of previous group link prediction methods suffer from missing information and weak generalization. To this end, we propose BiGLP, a novel group link prediction method based on graph neural networks (GNNs), to infer links between individuals and groups in bipartite graphs. To model intra-group relationships, we first design a new GNN with sampling strategy to learn representations of individuals by capturing neighborhood information. Moreover, we extract features from neighborhood of groups to accurately model inter-group relationships. From a new perspective that combining intra-group and inter-group relationships, BiGLP finally obtains representations of groups and predicts the targets based on group vectors. Experimental results on four datasets show that, in three evaluation metrics, BiGLP obtains average gains of 2.8%, 8.2% and 3.9%.