Fairness Aware Regularization on a Learning-to-Rank Recommender System for Controlling Popularity Bias in E-Commerce Domain

Abstract

Recommender systems aim to recommend items for certain users based on rating prediction, user preferences, or other parameters. Along with their advantages in guiding users to appropriate items, recommender systems potentially experience a popularity bias. It is a condition where popular items will often be recommended, while non-popular items will never be recommended. Controlling popularity bias is necessary to introduce new, but appropriate, items. It can be done by making a fair proportion of popular and non-popular items to be recommended to users. One of the methods is by applying fairness aware regularization. This method can increase the proportion of medium-tail recommendations of non-popular items, without sacrificing ranking performance on the recommender system. We combined this method with a learning-to-rank algorithm and applied it in a recommender system for e-commerce. Based on the testing results by using a dataset of Amazon review, the method is able to increase the proportion of medium-tail recommendation of non-popular items, up to 54.1%, on the measurement of Average Percentage of Tail (APT) and gain 0.281 on the measurement of Normalized Discounted Cumulative Gain (NDCG) score. In addition, the method is also able to introduce different medium-tail non-popular items to users.