A Novel Hybrid Item-Based Similarity Method to Mitigate the Effects of Data Sparsity in Multi-Criteria Collaborative Filtering

Data sparsity presents a significant challenge for Recommendation Systems, particularly in neighborhood-based approaches that rely on co-ratings to compute similarity. As co-ratings decrease, these methods often struggle to generate accurate recommendations. Addressing the persistent challenge of data sparsity, a novel similarity method, hybrid item-based, for multi-criteria collaborative filtering is proposed in this study. The proposed approach introduces a local similarity measure that combines a modified Jaccard similarity with a distance-based method, enhancing item-to-item similarity calculations. Additionally, it incorporates α-divergence as a global similarity measure, leveraging the probability density distribution of ratings to compute similarities even when user overlap is minimal or absent. This dual approach allows our method to produce reliable similarity measures in sparse data environments. The results of experimental evaluations conducted on multiple datasets illustrate that the proposed approach demonstrates a superior performance than existing traditional and hybrid approaches in terms of accuracy and achieves notable improvements in recall and F1-score. Furthermore, the proposed method exhibits high diversity and competitive novelty, striking an optimal balance that enhances user engagement. These results underline the proposed method’s potential as a comprehensive solution, offering improved accuracy, diversity, and relevance, particularly in sparse data contexts.