Multi-space interaction learning for disentangled knowledge-aware recommendation

Abstract

Knowledge graphs are becoming increasingly valuable as auxiliary information in current recommendation systems. User–item interaction modeling (within user–item interaction graphs) and knowledge association modeling (within knowledge graphs) play a vital role in representation learning. Most existing approaches are usually limited to modeling user–item interactions and knowledge associations in one space. However, they often exhibit distinctly non-Euclidean anatomy with diverse geometric features (e.g., trees and cyclic structures). In addition, models based on a single space cannot adequately accommodate multiple geometric patterns and ignore interaction factors between mixed geometries. To address these two issues, we utilize multi-space interactive learning for disentangled knowledge-aware recommendation (MKRec), which learns disentangled representations of users and items in multiple spaces, including hyperbolic, spherical, and Euclidean spaces, to preserve the geometric patterns of the data. Furthermore, to consider the interaction between mixed geometric shapes, we introduce an interactive learning mechanism along with a novel multi-space feature fusion method. Based on the empirical results, the evaluation of four datasets demonstrates the excellent efficacy of MKRec. In particular, our findings confirmed that the MKRec algorithm achieved significant results in disentangling user–item interactions and knowledge associations.