Enhancing Heterogeneous Knowledge Graph Completion with a Novel GAT-based Approach

TL;DR

This work addresses knowledge graph completion in heterogeneous KGs, where overfitting and entangled neighbor signals hinder GAT-based methods. It introduces GATH, an encoder-decoder framework with two attention streams: an entity-specific attention network and an entity-relation joint attention network, plus a reduced-feature encoding and weight-sharing scheme to curb parameter explosion. The model employs a ConvE-based decoder and a relation-aware feature transformation to improve tail/head predictions, achieving state-of-the-art results on FB15K-237 and WN18RR with notable gains in Hits@1, Hits@10, and MRR. This approach advances KGC for complex, multi-relational graphs and holds practical implications for more accurate search and recommendations in large heterogeneous knowledge graphs, with avenues for future work in further reducing over-smoothing and extending to complex number representations.$

Abstract

Knowledge graphs (KGs) play a vital role in enhancing search results and recommendation systems. With the rapid increase in the size of the KGs, they are becoming inaccuracy and incomplete. This problem can be solved by the knowledge graph completion methods, of which graph attention network (GAT)-based methods stand out since their superior performance. However, existing GAT-based knowledge graph completion methods often suffer from overfitting issues when dealing with heterogeneous knowledge graphs, primarily due to the unbalanced number of samples. Additionally, these methods demonstrate poor performance in predicting the tail (head) entity that shares the same relation and head (tail) entity with others. To solve these problems, we propose GATH, a novel GAT-based method designed for Heterogeneous KGs. GATH incorporates two separate attention network modules that work synergistically to predict the missing entities. We also introduce novel encoding and feature transformation approaches, enabling the robust performance of GATH in scenarios with imbalanced samples. Comprehensive experiments are conducted to evaluate the GATH's performance. Compared with the existing SOTA GAT-based model on Hits@10 and MRR metrics, our model improves performance by 5.2% and 5.2% on the FB15K-237 dataset, and by 4.5% and 14.6% on the WN18RR dataset, respectively.

Figures (3)

• Figure 1: The examples of heterogeneous knowledge graphs, where the circle represents the entity, labeled as $E_x$ and $x$ is the number. The arrows indicate the relations and different color arrows indicate different types of relations. The relation is noted as $r_x$.
• Figure 2: Entity information is represented by a combination of information in different scenarios. In the feature vector, different dimensions correspond to different scenes. A relation is a concrete representation of a scene.
• Figure 3: Framework of GATH. GATH consists of two parts: encoder and decoder. In the entity-specific attention network module, an entity's attention score to its neighbors depends on the neighbors themselves. In the entity-relation joint attention network module, an entity's attention score to its neighbors depends on the relation type. Similarly, the aggregation module aggregates information from the entity-specific attention network module and the entity-relation joint attention network module.