Enhancing missing facts inference in knowledge graph using triplet subgraph attention embeddings

Abstract

Due to the heterogeneous structure of the knowledge graph (KG), relationships between entities remain missing. However, optimal use of KG requires inference of missing fact triplet (entity-relation-entity). The fact inference predicts a missing relationship using an embedding approach in a supervised learning setup, representing entities and relationships in a low-dimensional vector space. Recent work uses attention-aware embeddings, but when applied directly to entire KG, attention mechanisms can be computationally expensive, especially for large graphs. The attention-based KG embedding model uses negative sampling, which can cause a gradient vanishing problem during learning. This paper proposes a novel triplet subgraph attention embedding (TSAE) model that combines a simplified graph attention mechanism with a neural network to learn embedding without negative sampling requirements. The attention layer processes the triplet-level subgraph entities to learn the central entity features by aggregating the neighbor’s features. A neural network processes attention-aware triplet entity features through hidden layers to compute the likelihood of relationship types between triplet entities. TSAE generates more fine-grained entity embeddings using simplified attention mechanism, reduces computational complexity, and offers interpretable embeddings. Experimental results on the benchmark data sets exhibit TSAE superiority over the baselines. The case study shows the efficacy of the model for the KG completion task.