Dual-Perspective Disentangled Multi-Intent Alignment for Enhanced Collaborative Filtering

Abstract: Disentangling user intents from implicit feedback has emerged as a promising strategy for enhancing both the accuracy and interpretability of recommendation systems. However, existing methods often model user and item intents independently and rely heavily on implicit structural signals, lacking explicit guidance to uncover the joint semantics that drive user-item interactions. To address these limitations, we propose DMICF, a dual-perspective collaborative filtering framework that unifies intent alignment, structural fusion, and discriminative training into a cohesive architecture. DMICF jointly encodes user-item graphs from both user and item views, leveraging cross-perspective structural signals to reinforce representation learning, especially under sparse or long-tail scenarios. A sub-intent alignment mechanism is introduced to uncover fine-grained semantic correspondences between users and items, enabling adaptive refinement of interaction representations. To enhance prediction quality, DMICF employs an intent-aware scoring module that aggregates compatibility signals across matched latent intents. Furthermore, a multi-negative softmax-based supervision strategy is incorporated to promote semantic disentanglement, encouraging alignment between relevant intents while suppressing spurious or entangled components. Extensive experiments confirm that DMICF consistently delivers robust performance across datasets with diverse interaction distributions. Qualitative analysis confirms that DMICF disentangles interaction intents and adaptively structures intent subspaces into semantically coherent clusters, enabling fine-grained personalization.