Adaptive fuzzy prototype learning with view-balanced multi-view embeddings for multi-behavior recommendation

Multi-behavior recommender systems(RS) leverages diverse user actions(e.g. click, cart, purchase) to model preferences more comprehensively, yet it often suffers from noisy behavior signals and heterogeneous user–item representations. In particular, integrating multiple behaviors enriches supervision but also introduces substantial interaction noise, while treating users and items homogeneously may lead to isolated embeddings and weak modeling of their latent associations. To address these challenges, we propose FMCL, which learns adaptive fuzzy prototypes from view-balanced multi-view embeddings and optimizes representations via contrastive learning for multi-behavior recommendation. FMCL first incorporates a context-aware feature fusion operator within a graph neural network to dynamically integrate implicit and explicit features from user- and item-centric perspectives, producing robust node representations. It then constructs three complementary graph views, namely user–user, user–item, and item–entity, and applies view-specific augmentations to generate informative yet noise-reduced contrastive views. Based on the original and augmented embeddings, FMCL introduces a view-balancing mechanism that adaptively weights cross-view discrepancies during fuzzy clustering, yielding stable prototype centroids and soft memberships. These fuzzy prototypes further guide multi-view contrastive learning, and the whole model is jointly optimized with a ranking objective to enhance recommendation accuracy and generalization. Extensive experiments on multiple public benchmarks demonstrate that FMCL consistently outperforms strong baselines, validating its effectiveness and robustness in supporting digital inclusion and inclusive innovation within multi-behavior recommendation scenarios.