InfoRank: Unbiased Learning-to-Rank via Conditional Mutual Information Minimization

Jiarui Jin, Zexue He, Mengyue Yang, Weinan Zhang, Yong Yu, Jun Wang, Julian McAuley

Keywords: Unbiased Learning-to-Rank, Conditional Mutual Information, Position Bias, Popularity Bias

TL;DR: Our main idea is to consolidate the impacts of position and popularity biases into a single observation factor, thereby providing a unified approach to addressing bias-related issues,

Abstract: Ranking items regarding individual user’s interests is a core technique of multiple downstream tasks such as recommender systems. Learning such a personalized ranker typically relies on the implicit feedbacks from users’ past click-through behaviors. However, collected feedbacks are biased toward previously highly-ranked items and directly learning from it would result in “rich-get-richer” phenomena. In this paper, we propose a simple yet sufficient unbiased learning-to-rank paradigm named InfoRank that can simultaneously mitigate the effects of both position and popularity biases. To this end, we first summarize the impacts of those biases into a single observation factor, therefore providing a unified treatment of the bias problem. We then minimize the mutual information between the observation estimation and the relevance estimation conditioned on the input features. By doing so, our relevance estimation can be proved to be free of bias. While implementation, we first adopt an attention mechanism to capture hidden correlations among user-item features and produce the estimations of observation and relevance. We then introduce a regularization term based on the conditional mutual information to encourage our relevance estimation to be conditionally independent of that of the observation. Experiments over three large-scale recommendation and search datasets demonstrate that InfoRank learns more accurate and unbiased ranking strategies from biased feedback compared to strong baselines under a variety of user browsing patterns.

Track: Search

Submission Guidelines Scope: Yes

Submission Guidelines Blind: Yes

Submission Guidelines Format: Yes

Submission Guidelines Limit: Yes

Submission Guidelines Authorship: Yes

Student Author: Yes

Submission Number: 262