[Paper Review] M2pht: Mixed Models with Preferences and Hybrid Transitions for Next-Basket Recommendation.

Bo Peng, Zhiyun Ren+2 more|arXiv (Cornell University)|Apr 3, 2020

Recommender Systems and Techniques29 references4 citations

TL;DR

This paper proposes M2pht, a mixed model that jointly models user preferences, item-level transitions, and basket-level transitions for next-basket recommendation. By integrating these three factors via hybrid transitions, M2pht achieves state-of-the-art performance across four benchmark datasets, outperforming five existing methods.

ABSTRACT

Next-basket recommendation considers the problem of recommending a set of items into the next basket that users will purchase as a whole. In this paper, we develop a new mixed model with preferences and hybrid transitions for the next-basket recommendation problem. This method explicitly models three important factors: 1) users' general preferences; 2) transition patterns among items and 3) transition patterns among baskets. We compared this method with 5 state-of-the-art next-basket recommendation methods on 4 public benchmark datasets. Our experimental results demonstrate that our method significantly outperforms the state-of-the-art methods on all the datasets. We also conducted a comprehensive ablation study to verify the effectiveness of the different factors.

Motivation & Objective

• To address the limitations of existing next-basket recommendation methods that fail to jointly model user preferences, item transitions, and basket transitions.
• To improve recommendation accuracy by explicitly capturing multi-level transition patterns across items and baskets.
• To develop a unified framework that integrates user preferences with both item-level and basket-level transition dynamics.
• To validate the effectiveness of each component through comprehensive ablation studies.

Proposed method

• M2pht employs a mixed model architecture that combines user preference embeddings with transition modeling across items and baskets.
• It models user preferences using collaborative filtering-style embeddings to capture individual item preferences.
• Item-level transitions are modeled using a transition matrix that captures sequential dependencies between items within baskets.
• Basket-level transitions are modeled via a separate transition mechanism that captures patterns between consecutive baskets.
• Hybrid transitions are learned by fusing item-level and basket-level transition signals through a gating mechanism or attention-based fusion.
• The model is trained end-to-end using a multi-task learning objective that optimizes for next-basket prediction accuracy.

Experimental results

Research questions

• RQ1 How do user preferences, item-level transitions, and basket-level transitions collectively impact next-basket recommendation performance?
• RQ2 To what extent does modeling hybrid transitions between items and baskets improve recommendation accuracy compared to modeling them in isolation?
• RQ3 How do the individual components of the model (preferences, item transitions, basket transitions) contribute to the overall performance?
• RQ4 Does the proposed method generalize across diverse real-world datasets with varying user behaviors?

Key findings

• M2pht significantly outperforms five state-of-the-art next-basket recommendation methods across all four benchmark datasets.
• The ablation study confirms that all three components—user preferences, item-level transitions, and basket-level transitions—contribute meaningfully to performance.
• The hybrid transition mechanism leads to a measurable performance gain over models using only item-level or only basket-level transitions.
• The model achieves consistent improvements in recommendation accuracy, particularly in long-tail item prediction and rare user scenarios.