Modeling Multi-Intent Basket Sequences for Next-Basket Recommendation

Abstract

Recommendation systems have a preponderance in assisting customers to save time by suggesting relevant options. With this convenience, a customer may purchase multiple items in a browsing session, referred to as an item basket. The notion of basket manifests his underlying preference of multiple implicit intentions, which becomes more sophisticated once considering the basket sequence of his chronological intersession list. With the objective of modeling basket sequences, most of previous methods hypothesize a homogeneous intention in each basket. The exploitation on multi-intent basket sequences for the recommendation task becomes an emerging demand. In this work, we present a novel framework named MIBS to model multi-intent basket sequences to recommend next basket of relevant items. Given a user's basket sequence, each basket is encoded via aggregating the item-item correlation matrix with a latent intent parameter matrix to generate the respective basket representation. This representation is later fed into a LSTM layer to infer the sequential encoding, which is also combined with the correlation matrix and the multi-intent matrix to produce item scores. The top-K items with the highest scores are employed to form the next-basket suggestion. Comprehensive experiments on three publicly-available datasets demonstrate the superiority of MIBS compared against state-of-the-art baselines for the next-basket recommendation task.