1 Introduction
Recent years have witnessed flourishing publications on recommendation, most of which aim at inventing machine learning models to fit users’ historical behavior data [1]. However, the observation data usually exhibits severe popularity bias, i.e., the distribution over items is quite imbalanced and even long-tailed. Such skewed distribution may be caused by the users’ conformity, deviating from reflecting users’ true preference. As a crucial factor for users’ decision-making, conformity describes the tendency that user behaves following the group. In a typical recommender system, a user may click an item simply because he finds the item clicked by many other users, rather than based on his own judgement. As a result, recommendation model trained on such biased data would yield unexpected results, e.g., capturing skewed user preference and amplifying the long-tail effect. Given the wide existence of popularity bias and its negative impact on recommendation, we cannot emphasize too much the importance of tackling popularity bias.