Multi-feature Subgraph Fusion with Text Knowledge on Citation Link Prediction

We propose multi-feature subgraph fusion neural networks to predict the citation links. We aim to refine the sparsity of the subgraph feature of citation links among articles. The proposed model fuses four features: subgraph of social network metrics, metadata article info, Word2Vec, and tf-IDF of the articles. Basically, we focus on the edge list feature level instead of the graph level since we can avoid the heavy computation for the adjacency matrix. However, we may neglect the similarity between articles in the text domain. Henceforth, we fuse with text knowledge from the articles, such as the corpus embedding and metadata of the articles. The proposed model was evaluated on a public dataset for link prediction. We can enjoy the proposed model performances by showing the ablation study on the fusion feature(s).