Title:Long-Tail Session-based Recommendation from Calibration

Abstract:Recent years have witnessed progress in precisely predicting the user's next behavior in session-based recommendation scenarios. However, recommending long tail items which are proved important to recommender systems is rarely investigated in existing work. To handle this problem, we incorporate the calibration for long tail session-based recommendation, which aims to align the proportion of tail items of recommendation list with the proportion in the session. To do this, we design a calibration framework to make the model aware of the popularity distribution of it recommendation list, and calibrate the result according to the ongoing session. Meanwhile a separate training and prediction strategy is applied to deal with the imbalance problem caused by popularity bias. Experiments on benchmark datasets shows that our model can both achieve the competitive accuracy of recommendation and provide much more tail items.