Dual stream fusion link prediction for sparse graph based on variational graph autoencoder and pairwise learning

Abstract

Recently, link prediction methods based graph neural networks have garnered significant attention and achieved great success on large datasets. However, existing methods usually rely on explicit graph structures, which is hard to obtain in sparse graphs. In addition, the incomplete graph data used for model training may lead to distribution shift between training and testing sets. To address these issues, this paper proposes a novel link prediction method for sparse graphs based on variational graph autoencoder and pairwise learning. By incorporating noise perturbation variational autoencoders, the proposed method can enhance robustness during sparse graph training. Instead of relying on explicit graph features, we reconstruct the original adjacency matrix by disturbing node feature mean encoding or variance encoding. To mitigate the impact of insufficient topological information, we introduce pairwise learning scheme, which obtains pairwise edges through negative sampling and iteratively optimize the positive and negative complementary probability adjacency matrix. Furthermore, we integrate the probability adjacency matrix and node similarity prediction based on message passing networks into a dual-stream framework to predict unknown links. Experimental results on multiple sparse networks demonstrate the superior link prediction performance of our proposed method over baseline approaches. Our method improves AUC from 0.3% to 1.5% and Precision from 1.4% to 5.2% across seven datasets.