[Paper Review] M2: Mixed Models with Preferences, Popularities and Transitions for Next-Basket Recommendation

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

Recommender Systems and Techniques26 references5 citations

TL;DR

This paper proposes M², a mixed model for next-basket recommendation that integrates users' general preferences, item global popularities, and item transition patterns using a novel encoder-decoder framework (ed-Trans). Unlike RNN-based methods, M² models individual item transitions more effectively, achieving up to 22.1% improvement over state-of-the-art methods across four benchmark datasets in next-basket prediction tasks.

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 novel mixed model with preferences, popularities and transitions (M2) for the next-basket recommendation. This method models three important factors in next-basket generation process: 1) users' general preferences, 2) items' global popularities and 3) transition patterns among items. Unlike existing recurrent neural network-based approaches, M2 does not use the complicated networks to model the transitions among items, or generate embeddings for users. Instead, it has a simple encoder-decoder based approach (ed-Trans) to better model the transition patterns among items. We compared M2 with different combinations of the factors with 5 state-of-the-art next-basket recommendation methods on 4 public benchmark datasets in recommending the first, second and third next basket. Our experimental results demonstrate that M2 significantly outperforms the state-of-the-art methods on all the datasets in all the tasks, with an improvement of up to 22.1%. In addition, our ablation study demonstrates that the ed-Trans is more effective than recurrent neural networks in terms of the recommendation performance. We also have a thorough discussion on various experimental protocols and evaluation metrics for next-basket recommendation evaluation.

Motivation & Objective

• To address the limitations of existing RNN-based next-basket recommendation methods that poorly model individual item transitions and user preferences due to data sparsity.
• To develop a unified framework that explicitly models three key factors: user preferences, item popularities, and item transition patterns.
• To evaluate the effectiveness of a simple, efficient encoder-decoder architecture (ed-Trans) for modeling item-level transitions compared to complex RNNs.
• To provide a comprehensive analysis of evaluation protocols and metrics for next-basket recommendation, advocating for NDCG and recall-based metrics over precision@k.
• To identify the complementarity and relative importance of each factor in improving recommendation accuracy.

Proposed method

• M² integrates three components: user general preferences (learned via collaborative filtering), item global popularities (frequency-based), and item transition patterns (modeled via ed-Trans).
• The ed-Trans framework uses an encoder-decoder architecture to model sequential transitions between individual items across baskets, capturing short-term dependencies without RNNs.
• Three variants are proposed: M²-p² (preferences + popularities), M²-gp² (personalized weights for preferences and popularities), and M²-gp²t (full model with all three factors).
• Weights for combining factors are learned via a weighted fusion mechanism, with personalized weights in M²-gp² to adapt to individual user behaviors.
• The model is trained end-to-end using negative sampling and stochastic gradient descent, optimizing for ranking metrics like recall@k and NDCG@k.
• Ablation studies isolate the contribution of each component, and visualization of learned item embeddings confirms ed-Trans captures ground-truth transition patterns.

Experimental results

Research questions

• RQ1 How do user preferences, item popularities, and item transition patterns collectively influence next-basket recommendation performance?
• RQ2 Can a simple encoder-decoder framework (ed-Trans) outperform complex RNN-based models in modeling item-level transitions for next-basket prediction?
• RQ3 What is the relative contribution of each factor (preferences, popularity, transitions) to overall recommendation accuracy?
• RQ4 Which evaluation metrics—precision@k, recall@k, or NDCG@k—are most appropriate and reliable for next-basket recommendation tasks?
• RQ5 How do different experimental protocols and data sparsity levels affect the robustness and generalization of next-basket recommendation models?

Key findings

• M²-gp²t, the full model combining preferences, popularities, and transitions, achieves the highest performance, outperforming all 5 state-of-the-art methods by up to 22.1% in NDCG@k and 19.0% in recall@5.
• The ed-Trans framework significantly outperforms RNN-based transition modeling, achieving up to 20.4% improvement in recall@5 on benchmark datasets.
• Ablation studies confirm that user preferences are the most critical factor, with transitions and popularities providing complementary improvements.
• The learned item embedding matrix from ed-Trans effectively captures ground-truth transition patterns, validating its representational quality.
• Precision@k is shown to be unreliable due to non-monotonicity and bias toward basket size, while recall@k and NDCG@k are more robust and informative for evaluation.
• The model demonstrates strong generalization across four public datasets, including retail, beauty, and grocery benchmarks, under varying data sparsity levels.