Exploiting uninteresting items for effective graph-based one-class collaborative filtering

The goal of recommender systems is to identify the items appealing to a target user by analyzing her/his past preferences. Collaborative filtering is one of the most popular recommendation methods that use the similarity between users' past behaviors such as explicit user ratings (i.e., multi-class setting) or implicit click logs (i.e., one-class setting). Graph-theoretic one-class collaborative filtering (gOCCF) has been successful in dealing with sparse datasets in one-class settings (e.g., clicked or bookmarked). In this paper, we point out the problem that gOCCF requires long processing time compared to existing OCCF methods. To overcome the limitation of the original gOCCF, we propose a new gOCCF approach based on signed random walk with restart (SRWR). Using SRWR, the proposed approach accurately and efficiently captures users' preferences by analyzing not only the positive preferences from rated items but also the negative preferences from uninteresting items. We also perform an in-depth analysis to further understand the effect of employing uninteresting items in OCCF. Toward this end, we employ the following well-known graph properties: (1) effective diameter, (2) number of reachable pairs, (3) number of nodes in the largest connected component, (4) clustering coefficient, (5) singular values, and (6) signed butterfly. From this comprehensive analysis, we demonstrate that signed graphs with uninteresting items have properties similar to real-life signed graphs. Lastly, through extensive experiments using real-life datasets, we verify that the proposed approach improves the accuracy and decreases the processing time of the original gOCCF.

Automated summary

This paper proposes a new gOCCF approach based on signed random walk with restart (SRWR) to improve the accuracy and efficiency of capturing users' preferences on items. Through in-depth analysis of the effect of employing uninteresting items, it demonstrates that this method can enhance accuracy and decrease processing time in real-life datasets.