Integrating neighborhood-based method and matrix completion for enhanced recommender systems

Recommender systems deliver personalized item suggestions to users by predicting their likelihood of purchase action for specific items. One of the most challenging issues in recommender systems is rating sparsity, which occurs due to users’ insufficient initial information on preferences. To address this issue, we propose a novel algorithm that dynamically integrates the neighborhood-based method with matrix completion, effectively improving prediction accuracy even under severe data sparsity. Specifically, we identify top-N neighbors using cosine similarity to enhance prediction accuracy during the iterative completion of the rating matrix. Furthermore, traditional matrix completion algorithms often suffer from high computational latency when scaling to large datasets or handling frequent new user enrollments. To resolve this computational bottleneck, we propose a computationally efficient modification of the algorithm by reusing previous computations rather than computing the algorithm from scratch. Experimental results demonstrate that the proposed algorithm significantly improves the prediction accuracy of new users’ preferences compared to existing baseline algorithms while maintaining fast execution speeds.