A Graph Embedding Approach for Link Prediction via Triadic Closure Based Direct Aggregation and Weighted Concatenation

Yahui Chai, Xiaobin Rui, Jie Yang, Philip S. Yu, Zhixiao Wang

Abstract: Graph embedding based on deep learning is an effective approach for link prediction. However, there still remain some unsolved problems. Firstly, existing methods iteratively aggregate node embeddings from the neighborhood, which cannot retain the global structural information at lower node aggregation cost. Secondly, existing methods can hardly retain rich local structural information for edge embeddings obtained from node embeddings by Hadamard product, summation, or direct concatenation. To tackle these challenges, this paper proposes a novel graph embedding approach for link prediction via triadic closure based direct aggregation and weighted concatenation. To aggregate neighborhood information with high efficiency, our proposed approach directly aggregates multi-order neighbors to the central node and utilizes the triadic closure structure to assign different aggregating weights. To well retain the local structural information, our proposed approach generates edge embeddings through weighted summation. Extensive experiments demonstrate that the triadic closure based direct aggregation and weighted concatenation enable our proposed approach to efficiently learn more accurate embeddings for link prediction, outperforming state-of-the-art methods.