Long-Tail Session-based Recommendation from Calibration

Abstract

Accurate prediction in session-based recommendation has progressed, but a few studies focused on skewed recommendation lists caused by popularity bias. Existing models on mitigating the popularity bias attempted to reduce the over-concentration on popular items by simply amplifying scores of less popular items. However, they normally ignore the users' different preferences towards tail items. To this end, we incorporate calibration, where calibrated recommendations reflect the user's interests with appropriate proportions, to mitigate the popularity bias from the user's perspective. Concretely, we add a calibration module to predict the ratio of tail items in the recommendation list from the session representation, and align it to the ongoing session. Meanwhile, we propose a two-stage curriculum training strategy to improve the prediction in the calibration module. Experiments on benchmark datasets show that our model can both achieve the competitive accuracy of recommendation and provide more tail items.