COLLABORATIVE FILTERING BASED RESOURCE RECOMMENDATION WITH USER-EXPERIENCE DRIVEN EVALUATION METRICS

Abstract

Digital libraries and large-scale information repositories increasingly rely on Collaborative Filtering (CF) techniques to help users navigate vast collections of resources. Although CF has been widely studied and applied for more than a decade, prevailing algorithms still exhibit limitations that negatively influence the overall user experience. This article empirically shows that several popular CF algorithms, while effective in traditional accuracy-based evaluations, fail to model user interaction with library resources in a realistic and user-centric manner. To address these shortcomings, we propose a new Belief Distribution Algorithm, designed to enhance the quality of resource recommendations in library information systems. Unlike conventional point-prediction approaches, the proposed model generates belief distributions across possible rating values, enabling a more nuanced and comprehensive representation of user preferences. The algorithm maintains the strengths of classical nearest-neighbor methods while significantly improving predictive depth and interpretability. Furthermore, our findings reveal that the widespread reliance on mean absolute error (MAE) has masked critical weaknesses in existing algorithms. As an alternative, we introduce a user-experience-driven Precision metric, specifically adapted for library resource recommendation scenarios. The proposed metric provides a more accurate assessment of how well recommendations align with actual user needs and behaviors. Overall, this study advances CF research by presenting a more realistic, user-centered approach to evaluating and improving recommendation performance in digital library environments.