Original scientific paper

https://doi.org/10.17559/TV-20250428002631

Enhancing Recommendation Accuracy of Item-Based Collaborative Filtering under Data Sparsity

Tao Zhang ; 1) Information School, Yunnan University of Finance and Economics, Kunming, China 2) Engineering Research Center of Cyberspace, Kunming, China
Jiaming Pi ; Graduate school, Yunnan University of Finance and Economics, Kunming, China *
Kun Zhao ; Information School, Yunnan University of Finance and Economics, Kunming, China

* Corresponding author.

Abstract

Data sparsity remains a major challenge for collaborative filtering (CF) systems, where enhancing recommendation accuracy with limited data is crucial. Traditional CF methods depend on common ratings for similarity computation but often ignore users' bounded rationality in neighbor selection and rating behavior. To overcome these limitations, we propose an Expected Utility-based Collaborative Filtering (EUCF) method with three key contributions: (1) a positioning function computing item expected ratings to capture global patterns, (2) a dynamic similarity criterion incorporating item rating frequency and user rating saturation for adaptive neighbor selection (threshold ε), and (3) an average deviation correction model addressing imperfect ratings. Experiments on MovieLens-1M and Netflix show EUCF outperforms KNN-IBCF, improving MAE by 3.56% and 4.17%, and MSE by 6.02% and 6.45%, respectively. Moreover, EUCF boosts recommendation coverage by 9.96% and 11%, mitigating data sparsity and long-tail challenges. With reduced computational complexity, EUCF ensures better scalability for real-world deployment. The results demonstrate its effectiveness in balancing accuracy, efficiency, and robustness under sparse data conditions.

Keywords

bounded rationality; collaborative filtering; data sparsity; expected utility; recommendation accuracy

Hrčak ID:

342629

URI

https://hrcak.srce.hr/342629

Publication date:

31.12.2025.

Visits: 455 *