Date of Award

2023

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Multi-task learning, Knowledge graph completion, Tasks’ performance, Relation patterns

Supervisor

Z. Kobti

Supervisor

K. Selvarajah

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Advances in multi-task learning (MTL) models have improved the performance and explainability of recommender systems (RS) by jointly learning the recommendation and knowledge graph completion (KGC) tasks. Recent studies have established that considering the incomplete nature of knowledge graphs (KG) can further enhance the performance of RS. However, most existing MTL models depend on translation-based knowledge graph embedding (KGE) methods for KGC, which cannot capture various relation patterns, including composition relations that are prevalent in real-world KG.

To address this limitation, this thesis proposes a new MTL model, named rotational knowledge-enhanced translation-based user preference (RKTUP). RKTUP enhances the KGC task by incorporating rotational-based KGE techniques (RotatE or HRotatE) to model and infer diverse relation patterns. These relation patterns include symmetry/asymmetry, composition, and inversion. RKTUP is an advanced variant of the knowledge-enhanced translation-based user preference (KTUP) MTL model, which provides interpretations of its recommendations.

The experimental results demonstrate that RKTUP outperforms existing methods and achieves state-of-the-art performance on both recommendation and KGC tasks. Specifically, it shows a 13.7% and 11.6% improvement in F1 score for recommendations on DBbook2014 and MovieLens-1m, respectively, and a 12.8% and 13.6% increase in hit ratio for KGC on the same datasets, respectively.

The use of RotatE improves the two tasks’ performance, while HRotatE enhances the two tasks’ performance and the model’s efficiency.

Share

COinS