Path-KGE: Preference-Aware Knowledge Graph Embedding with Path Semantics for Link Prediction

Abstract

Knowledge graph embedding (KGE) aims to transfer entities and relations into low-dimensional vector space while preserving underlying semantics in knowledge graph (KG). Some existing models directly learn embeddings based on original triples, while others integrate external information to enhance the limited semantic information in KG. However, most of them ignore latent user preferences in their interactions with real-world KG applications, not even considering the relational semantics implied in historical interaction data, which are practical and essential for many downstream tasks like link prediction. To address these issues, we propose a novel preference-aware knowledge graph embedding with path semantics model Path-KGE, which learns semantic information of relation paths and integrates it into embedding process. First, we mine multi-hop relations with frequent and temporal characteristics as semantic relation paths in user interaction data, to obtain implicit user preference features. Second, we design a path importance function to distinguish the semantic impacts of user preferences on distinct relation paths. Finally, we utilize weight scores as long-term constraints to learn preference-aware embeddings, so that the representation vectors of entity pairs connected by preferred relation paths can remain close distances in vector space as in reality. The experimental results show that our model outperforms state-of-the-art translation-based models on link prediction and triple classification tasks.