Two-stage system using item features for next-item recommendation

Abstract

• We address next-item recommendation using user transactions and item features. • Our item embedding captures transaction-based similarity and item features. • The candidate items are generated by predicting next-item features. • Our model addresses the cold-start problem of item. The next-item recommendation task aims to predict the item that a user is likely to interact with next, given the previous transactions of the user. In this paper, we explore the use of discrete item features to improve the performance of a next-item recommendation system. We design a hybrid embedding method based on user transactions and item features to represent items that can capture item-item co-occurrence as well as item features. We propose a two-stage model for the task of next-item recommendation. In the first stage, features of the next item are predicted, based on which candidate items are generated. In the second stage, the candidate items are ranked, and the top item is recommended to the users. Our model is capable of recommending new items since the candidate items are selected based on item features. We evaluate our model on three different datasets, including two public datasets and show that our model achieves significant improvement compared with several state-of-the-art next-item recommendation models.