Mitigating sparsity using Bhattacharyya Coefficient and items’ categorical attributes: improving the performance of collaborative filtering based recommendation systems

摘要

Collaborative filtering has been the most popular and effective recommendation technique to predict ratings using similar users or items. But in a sparse dataset, due to fewer co-rated items, the traditional similarity measures fail to compute the similarity between a pair of users. This influences the predicted rating negatively, which results in degraded recommendation performance. Similarity calculation using Bhattacharya Coefficient can be a more judicious approach because it works well with few or no co-rated items between a pair of users. However, Bhattacharya Coefficient also fails to compute the similarity between a pair of users when co-rated items are zero and the rating vector of items are disjoint. In this paper, we propose a novel approach to address the limitation of the Bhattacharya Coefficient with improved rating prediction accuracy in collaborative filtering. Instead of using only user ratings, to have more rating prediction accuracy, we use categorical attributes of rated items in findings of k-nearest neighbors. The performance of the proposed approach is evaluated on the collected datasets of MovieLens and LDOS-CoMoDa and compared with recent approaches. The comparative results corroborate the anticipated performance of the proposed approach.