SQL-Rank++: A Novel Listwise Approach for Collaborative Ranking with Implicit Feedback

Abstract

Collaborative ranking with implicit feedback such as users' clicks is an important recommendation problem in various real-world applications. Most existing approaches are developed based on some pointwise or pairwise preference assumptions, although the listwise assumption is widely accepted as a better alternative due to its consistency with the final delivery result. In this paper, we first identify two fundamental limitations of the most current collaborative listwise approaches, in which their modeling is based on the Plackett-Luce probability. They are too strict and too weak relative preference comparison between the items with the same feedback and between the items with different feedback, respectively. As a response, we propose a novel and improved listwise approach called SQL-Rank++, which is able to learn the user preferences more accurately by leveraging some specifically constructed auxiliary lists, including some positive lists and some negative lists. Specifically, the positive lists have as much semantic consistency as the original list as possible, while the negative lists are the opposite. To construct these auxiliary lists, we design a self-based sampling strategy and a user similarity-based one. Finally, we have four variants of our SQL-Rank++ with different combinations of the auxiliary lists. We then conduct extensive experiments on four public datasets, and find that our SQL-Rank++ achieves very promising performance in comparison with several pointwise, pairwise and listwise approaches. We also study the influence of the two sampling strategies and the key components in our SQL-Rank++.