Improving Recommendation Accuracy using Cross-domain Similarity

Abstract

The accuracy of collaborative filtering based recommendation system depends on sufficient rating information of the target user and his/her neighbors for similar items. In a real scenario where a huge number of users and items exist, there may be a possibility of very less (high sparsity) or no (cold start problem) rating information in the rating dataset, which degrades the recommendation accuracy significantly. This opens up a scope for improvement in the prediction accuracy of the collaborative filtering based recommender system. In this paper, if the target user does not find k similar neighbors in a particular domain, the proposed algorithm utilizes the rating information of that user from other domains, if available. This not only reduces the sparsity in the rating dataset but also solves the cold start problem. Additionally, the modified similarity measure of the proposed approach not only considers the high ratings but also the low ratings which makes the recommendation more personalised. Overall, the proposed approach improves the recommendation accuracy to a great extent, which has been been evident from the accuracy measures (e.g., MAE and RMSE) of the comparative analysis of the proposed algorithm and other prevalent collaborative filtering methods, investigated on the Amazon dataset.