Recommender systems provide personalized recommendations by modeling user-item interactions, where disentangling users’ intents is critical for improving recommendation accuracy. While existing intent modeling methods aim to capture fine-grained intent representations, they face two challenges: 1) Neglecting the influence of social semantics on modeling fine-grained intents; 2) Implicit data sparsity and intent redundancy limiting intent characterization. To tackle these challenges, we propose a Social-Semantic Enhanced Dual-Intent Hypergraph Collaborative Filtering (SDIHGCF) model. Specifically, SDIHGCF constructs hypergraph structures to preserve social semantics among users, items, and groups. It encodes features from both social and interest perspectives to achieve user and item representations that integrate individual intent, which signifies private preferences, and collective intent, which denotes overall awareness. To mitigate data sparsity and intent redundancy, where one intent can be represented by others, we use graph contrastive regularization to enforce consistency among users, items, intents, and interactions. Additionally, a bidirectional contrastive learning loss is proposed to enhance intent alignment. Experiments on four datasets demonstrate that SDIHGCF outperforms existing methods, offering novel insights into fine-grained intent modeling. • Hypergraphs capture social semantics information. • Social-Semantic modeling enables disentanglement of dual intents. • Bidirectional contrastive learning mitigates sparsity and intent redundancy. • Outperforming SOTA methods on four datasets via dual-intent modeling.

Knowledge graphs, with their strong knowledge association capabilities, are widely applied in recommender systems. Knowledge-aware recommendation techniques tend to develop end-to-end models based on graph neural networks. While alleviating the long-tail problem of relations, existing models still face knowledge overload caused by the complex information (redundant edges and noise interference) and semantic bias caused by dependence on interaction data (interaction domination problem). To address these issues, we propose the Edge Pruning and Contrastive Learning Driven Recommendation via Virtual Relation Knowledge Graph (VRKG-EPCL) model. Specifically, it designs a two-layer knowledge graph pruning mechanism (random pruning and adaptive pruning) to construct a Virtual Relation Knowledge Graph, reducing noise and redundant connections in the knowledge graph; strengthens the modeling of high-order relation paths through the local weighted smoothing mechanism; and incorporates the contrastive learning strategy to enhance the discriminative ability of node representations in the feature space, effectively alleviating feature collapse caused by interaction domination. Experiments on relevant datasets show that the VRKG-EPCL model outperforms existing methods in performance, demonstrating its effectiveness.