RelpNet: Relation-based Link Prediction Neural Network

Node-based link prediction methods have occupied a dominant position in the graph link prediction task. These methods commonly aggregate node features from the subgraph to generate the potential link representation. However, in constructing subgraphs, these methods extract each node's local neighborhood from the target node pair separately without considering the correlation between them and the whole node pair. As a result, many nodes in the subgraph may have little contribution to predicting the potential edge. Aggregating these node features will reduce the model's accuracy and efficiency. In addition, these methods indirectly represent the potential link by the node embeddings in the subgraph. We argue that this formalism is not the best choice for link prediction. In this paper, we propose a relation-based link prediction neural network named RelpNet, which aggregates edge features along the structural interactions between two target nodes and directly represents their relationship. RelpNet first extracts paths between the target node pair as structural interactions, which have strong correlations with the whole node pair and fewer nodes and edges than node-based methods' subgraph. To aggregate edge embeddings along the links between edges, we propose transforming the paths into a line graph. Then, the Tree-LSTM model is adopted to transfer and aggregate the node embeddings in the line graph as a comprehensive representation of the target node pair. We evaluate RelpNet on 7 benchmark datasets against 15 popular and state-of-the-art approaches, and the results demonstrate its significant superiority and high training efficiency.