A personalized clustering-based approach using open linked data for search space reduction in recommender systems

Recommender systems use information about the users' preferences to define relatedness scores towards items. Regardless of the method, a noticeable problem is that the system is required to compute scores for a large amount of unknown items in the database, even though these items may not be related to a determined user. In this manuscript, we propose a technique called search space reduction for recommender systems (SSR4Rec) that reduces the number of unknown pairs the recommender must process. As a pre-processing step, we cluster related items and assign only the closest group to each user, producing a reduced set of unknown pairs. The distance between items, and between clusters and users, is computed by comparing item representations and user profiles built based on attributes extracted from the Linked Open Data cloud. We assess the quality of SSR4Rec by applying it into two well-known RS and comparing the results against the same recommenders without our pre-processing step, as well as against other related baselines. Results show a significant improvement in both ranking accuracy and computational time.