An experimental study for neighbor selection in collaborative filtering

Abstract

Similarity is a most critical index in collaborative filteringbased recommdender systems, which suggest items to their customers by consulting most similar neighbors. Current popular similarity measures may mislead the user to unwanted items in certain cases, due to their inherent properties. This study suggests a novel idea to significantly decrease such occurrences by enforcing qualifying conditions to neighbors using some simple criteria, to make consultations for their ratings. From extensive experiments, the proposed idea is found to substantially improve prediction performance of collaborative filtering based on existing similarity measures. This result is noticeable considering that such improvements are achieved by simply consulting only those neighbors satisfying the given criteria, without adopting any sophisticated similarity measure.