DyHuCoG: A Dynamic Hypergraph Cooperative Game for Preference-aware Recommendation

Recommender systems must balance accuracy, diversity, and robustness to enhance the user experience and engagement.We present DyHuCoG (Dynamic Hypergraph Cooperative Game), a cooperative-game-enhanced hypergraph recommender that models user-item-context interactions as dynamic coalitions.A preference-aware Shapley estimator (Monte Carlo) quantifies the marginal utility of each interaction, which is then injected as dynamic hyperedge weights into a lightweight hypergraph neural network with an interaction-level-attention gate.This design enables the model to adaptively weight user, item, and context signals, capturing nuanced preferences that conventional collaborative filtering and graph-based models often ignore.Evaluated on the MovieLens-1M and Amazon-Book datasets, DyHuCoG consistently outperforms strong baselines, including matrix factorization (MF), neural collaborative filtering (NCF), Light Graph Convolutional Network (LightGCN), Dual Contrastive Learning for Recommendation (RecDCL), Hypergraph Contrastive Collaborative Filtering (HCCF) and Hypergraph Projection Enhanced Collaborative Filtering (HPCF).Relative to HPCF (state-of-the-art baseline), DyHuCoG achieves improvements in Normalized Discounted Cumulative Gain at cutoff 20 (NDCG@20) and recall at cutoff 20 (Recall@20) by +9.77%/+12.58%(MovieLens-1M) and +13.33%/+16.16%(Amazon-Book) and further increases coverage and diversity.Although Shapley estimation introduces additional computation, end-to-end training remains practical at approximately 1.78× the duration of HPCF.These results demonstrate that integrating preference-aware cooperative game attributions directly into hypergraph message passing yields a lightweight, training-stable module that jointly improves ranking quality, coverage, and diversity for large-scale recommendation tasks.