Hypernetwork Link Prediction Method Based on Fusion of Topology and Attribute Features

Abstract

Link prediction aims at predicting missing or potential links based on the known information of complex networks. Most existing methods focus on pairwise low-order relationships while ignoring the high-order interaction and the rich attribute information of entities in the actual network, leading to the low performance of the model in link prediction. To mine the cross-modality interactions between the high-order structure and attributes of the network, this paper proposes a hypernetwork link prediction method for fusion topology and attributes (TA-HLP). Firstly, a dual channel coder is employed for jointly learning the structural features and attribute features of nodes. In structural encoding, a node-level attention mechanism is designed to aggregate neighbor information to learn structural patterns effectively. In attribute encoding, the hypergraph is used to refine the attribute features. The high-order relationship between nodes and attributes is modeled based on the node-attribute-node feature update, which preserves the semantic information jointly reflected by nodes and attributes. Moreover, in the joint embedding, a hyperedge-level attention mechanism is introduced to capture nodes with different importance in the hyperedge. Extensive experiments on six data sets demonstrate that this method has achieved a more significant link prediction effect than the existing methods.