Autoencoder-Based Sequence Modeling for Next Basket Recommendation

Abstract

This paper presents a novel approach in Next Basket Recommendation (NBR), focusing on the subtask of Next Novel Basket Recommendation (NNBR) in addition to conventional NBR. NNBR aims to suggest new items to customers, diversifying their purchasing patterns in retail settings. Our approach combines a pre-trained autoencoder and a Transformer architecture to process users' purchase histories. The autoencoder transforms purchase histories into a series of basket embedding vectors, which are then learned through the Transformer architecture. Our method effectively addresses the challenges of implicit feedback data, such as data scarcity and uncertainty, common in real-world retail situations. The model demonstrates competitive performance in NNBR under data with high exploration rates by learning personalized sequential basket purchase patterns, offering implications for enhancing customer shopping experiences and broadening product diversity in retail environments.