Knowledge and Information Systems Взаимодействие человека и компьютера

A Jensen–Shannon divergence-based item similarity method for multi-criteria collaborative filtering

Demirelli Okkalioglu, Burcu
Springer Nature · 2026-10-30 · DOI 10.1007/s10115-026-02885-0

Аннотация

Multi-criteria collaborative filtering enables users to evaluate items across multiple dimensions, leading to more refined recommendations. However, traditional methods, including multidimensional distance-based approaches, rely heavily on common ratings between users or items. In sparse data scenarios, this reliance often results in reduced accuracy due to limited overlapping ratings. Despite improvements in similarity calculations, many existing approaches continue to struggle with data sparsity and fail to capture complex item relationships. To address this challenge, a novel item-based similarity method is introduced, utilizing Jensen–Shannon divergence within a multi-criteria collaborative filtering framework. By computing item similarities based on probability density values rather than common ratings, the approach demonstrates enhanced effectiveness in sparse datasets where traditional correlation-based techniques are limited. When overlapping users are available, their count is incorporated as a weighting factor to further refine similarity calculations. Experimental results show that the proposed methods achieve lower prediction error than existing methods in most evaluated configurations, particularly under sparse conditions, with the advantage narrowing as dataset density increases. Statistical significance analysis indicates the robustness of these improvements across different experimental configurations, while runtime evaluation demonstrates computational efficiency comparable to existing sparsity-aware methods. These outcomes confirm that the proposed methods substantially improve recommendation accuracy in sparse data environments.

Читать у издателя →