AttFGCN: A GCN-Based Method Using Attention Flow for Knowledge Graph Completion

Mei Yu, Fan Gao, Mankun Zhao, Tianyi Xu, Jiujiang Guo, Wenbin Zhang, Dejun Hou, Ruiguo Yu

Abstract: Graph convolutional networks (GCNs) are a series of competitive methods for knowledge graph completion (KGC). GCNs can effectively capture the implicit information by aggregating neighbor embeddings. However, existing GCNs typically set the neighbor selection range based on the number of convolutional layers, which is a fixed hyperparameter. This rigid approach cannot dynamically adjust the neighbor selection range based on central entities and queries during training. Regarding above limitation, this paper introduces AttFGCN, an innovative GCN-based method using Attention Flow. This model makes the attention weight propagate from central entity to different paths in a breadth-first strategy, and analogizes this process to the physical process of incompressible fluid flowing in a pipe. Through concept transfer, the semantic correlations between entities, queries, and relation paths are modeled as different physical quantities to design the attention flow formulas. AttFGCN can flexibly determine different ranges on different paths to preferentially select higher correlation neighbors for feature aggregation. Comprehensive experiments conducted on the standard benchmarks demonstrate that AttFGCN outperforms existing state-of-the-art models in terms of the accuracy, portability, and robustness.