A lightweight hierarchical graph convolutional model for knowledge graph representation learning

Graph convolutional networks (GCNs) have emerged as powerful tools for handling graph-structured data. Many knowledge graph embedding models leverage GCNs as encoders to learn the relationships between central entities and their neighbors, showing impressive performance in the knowledge graph completion task. However, the incorporation of GCNs increases the computational burden of the model. Moreover, due to the complex graph structure of knowledge graphs, treating all neighboring entities equally will cause the model to lose its ability to capture important information. To address these challenges, we present a lightweight hierarchical graph convolutional network (Light-HGCN). Light-HGCN removes feature transformation and selectively uses nonlinearity activation in standard GCNs to accelerate model convergence. Additionally, Light-HGCN introduces a hierarchical attention mechanism to determine the neighboring weights to capture complex graph structures. Light-HGCN achieves promising performance for the link prediction task across multiple benchmark datasets. Ablation experiments further illustrate the effectiveness of the hierarchical attention mechanism. The analysis of feature transformation and nonlinearity activation in GCNs on the KGC task indicates that lightweight GCNs can enhance computational efficiency while preserving promising performance.