Neuro-symbolic Recommender Systems

Abstract

Recommender systems are a fundamental component of modern digital platforms, guiding how users discover information, products, services, and media content. Recent advances in deep learning have substantially improved recommendation performance. However, state-of-the-art approaches rely on black-box models whose limited interpretability and transparency hinder insight into their decision-making. This doctoral research investigates how symbolic methods can be integrated into neural recommender systems to improve interpretability while maintaining competitive recommendation performance. To this end, we study three complementary neuro-symbolic paradigms: (i) a hybrid model based on the cognitive architecture ACT-R; (ii) a differentiable fuzzy-logic network that learns rules describing user behavior for news recommendation; (iii) a relational transformer over knowledge-enhanced interaction graphs that enables cross-domain few-shot recommendation with feature-level attribution through in-context graph sampling. Preliminary results show that these paradigms achieve competitive accuracy while providing interpretable recommendations and transparent decision logic, highlighting the potential of neuro-symbolic paradigms as a foundation for more explainable and trustworthy recommender systems.