Consensus-Driven Group Recommendation on Sparse Explicit Feedback: A Collaborative Filtering and Choquet-Borda Aggregation Framework
This work addresses the problem of group decision-making for users with diverse preferences in sparse data settings, representing an incremental advancement in group recommendation methods.
The paper tackled the challenge of achieving stable consensus in group recommender systems with sparse explicit feedback by proposing a hybrid framework combining collaborative filtering and fuzzy aggregation, which improved group-level consensus, satisfaction, and fairness in experiments on real-world datasets.
Group Recommender Systems (GRS) play an essential role in supporting collective decision-making among users with diverse and potentially conflicting preferences. However, achieving stable intra-group consensus becomes particularly challenging when only sparse userID-itemID-rating data are available and no demographic, contextual, or group-level information exists. This paper proposes a consensus-driven hybrid group recommendation framework that integrates neighborhood-based collaborative filtering with fuzzy aggregation to support agreement, fairness, and robustness under sparsity. A composite similarity measure, CBS (Combined Similarity), is derived from two enhanced similarity metrics introduced in prior work: a geometry-based measure that captures rating-pattern structure, and an uncertainty-aware measure that models belief, evidence, and disagreement in sparse co-rating contexts. This combination provides more stable estimation of missing ratings and supports consensus-oriented neighborhood construction. Candidate items are generated by merging per-user top-N predictions and further enriched using the Borda Count mechanism to mitigate skewed rating distributions and reinforce group-level agreement. Final group ratings are computed using the Choquet integral, which flexibly captures heterogeneous user influence while preserving fairness and supporting consensus formation. Experimental results on real-world datasets with different rating distributions show that the proposed method improves group-level consensus, satisfaction, and fairness, while maintaining a balanced level of novelty. Although the model does not rely on social information, its evaluation using trust-aware novelty measures indicates stable behavior in socially structured environments.