Beyond Positive Similarity Metrics: Leveraging Negative Co-Occurrence in Recommender Systems

Given the growing amount of online information and the lack of effective tools to extract relevant content for individuals, recommender systems (RSs) have emerged as highly efficient means to provide personalized recommendations to users. Collaborative filtering (CF) is a popular recommendation technique that suggests items to the target user based on the ratings provided by similar users in the system. However, the performance of CF, specifically memory-based CF, relies on choosing the convenient similarity metric. An optimal choice of the similarity metric will lead to better recommendations. Despite extensive studies on RSs in the past decade, research on similarity metrics in RSs has only recently gained attention. This study investigates 37 similarity metrics that incorporate negative co-occurrence for implicit ratings. We conduct a comparative analysis of these metrics against various real-world datasets using well-known performance measures. Our study provides insights into effective similarity metrics and their performance in relation to data characteristics. Specifically, our results reveal that while the Russel-Rao similarity metric consistently outperforms other similarity metrics across various datasets in terms of precision, there is no superior similarity metric across all datasets and performance metrics.