Logic Preference Fusion Reasoning on Recommendation

User preference interests and logical requirements synergistically indicate users’ desires, which influence the user’s behaviors. While most models extract user preferences from interaction records, they frequently neglect the user’s logical requirements, which are embedded in the logical relations between items and entities. Existing methods that account for user’s logical requirements employ neural networks to mimic logical operators, failing to explicitly model logical relations. Therefore, we present a Logic Preference Fusion Reasoning (LPFR) framework on recommendation, which employs a neuro-symbolic approach to explicitly extract user preferences and logical requirements as user desires. Specifically, we design neuro-symbolic operators to compute from both feature and symbol perspectives, merging the computational benefits of embedding with the reasoning advantages of logic. Subsequently, we introduce a logic and preference reasoning module to reason user preferences and logical requirements explicitly, obtaining preference, logic and desire queries answer sets. This module allows LPFR to reason about user desires at perceptual and cognitive aspects, respectively. To explore specific information within the three answer sets, a co-guided fusion gate is introduced to mine mutual relations and co-guide learning. Experimental results indicate that our model surpasses existing baseline models.