RKTUP Framework: Enhancing Recommender Systems with Compositional Relations in Knowledge Graphs

Abstract

Advances in joint recommendation and knowledge graph completion (KGC) learning have enhanced the performance and explainability of recommendations. Recent studies have established that taking the incomplete nature of knowledge graphs (KG) into consideration can lead to enhancements in recommender systems’ (RS) performance. The existing models depend on translation-based knowledge graph embedding (KGE) methods for KGC. They cannot capture various relation patterns, including composition relations, even though the composition relationships are prevalent in real-world KG. This study proposes a simple and effective approach to enhance the KGC task while training it with the RS. Our approach, rotational knowledge-enhanced translation-based user preference (RKTUP), is an advanced variant of the knowledge-enhanced translation-based user preference model (KTUP), an existing MTL model. To enhance KTUP, we use the rotational-based KGE techniques (RotatE or HRotatE) to model and infer various relation patterns, such as symmetry/asymmetry, composition, and inversion. Unlike earlier MTL models, RKTUP can model and infer diverse relation patterns while learning more robust representations of the entities and relations in the KGC task, leading to improved recommendations for users. Using RotatE improved the recommender system’s performance while using HRotatE enhanced the model’s efficiency. The experimental results reveal that RKTUP outperforms existing methods and achieves state-of-the-art performance on the recommendation and KGC tasks. Specifically, it shows a 13.7% and 11.6% improvement in the F1 score, as well as a 12.8% and 13.6% increase in the hit ratio on DBbook2014 and MovieLens-1m, respectively.