Correlation-Sensitive Next-Basket Recommendation

Duc-Trong Le, Hady W. Lauw, Yuan Fang

Abstract

Next-basket recommendation is a critical task in e-commerce and social media, where the goal is to predict a set of items a user will adopt simultaneously. Traditional recommendation systems often struggle with this task because they either treat items within a basket as independent or fail to account for long-term sequential patterns in user behavior. This paper introduces the Basket Sequence COrrelation Network (Beacon), a hierarchical neural model designed to capture both intra-basket correlations and inter-basket sequential dependencies. By explicitly modeling item "coalitions" and their evolution over time, Beacon provides more coherent and accurate recommendations. Experimental results on three real-world datasets—TaFeng, Delicious, and Foursquare—show that Beacon consistently outperforms state-of-the-art baselines in terms of accuracy and utility.

Introduction

The next-basket recommendation problem involves predicting a collection of items likely to be purchased or used together in the future based on a user's historical sequence of baskets. Unlike standard recommendation tasks, this requires understanding the synergy between items. For instance, a user buying salmon is likely to also buy wasabi and rice for sushi. Existing approaches often fall short by ignoring these intra-basket correlations or by focusing only on short-term transitions, missing the broader context of a user's preferences.

Comparison between independent and correlation-sensitive recommendation.

Methodology

The proposed Beacon model addresses these challenges through a three-stage hierarchical architecture designed to integrate correlative and sequential signals.

  1. Correlation-Sensitive Basket Encoder: This component represents each basket by considering the "self-importance" of individual items alongside their pairwise correlations. It utilizes a Laplacian-normalized correlation matrix $C$ derived from the training data. A learnable parameter $\eta$ is employed to prune weak, noisy correlations, ensuring the model focuses on meaningful item associations.
  2. Basket Sequence Encoder: To capture long-term dependencies, the encoded basket representations are processed by a Long Short-Term Memory (LSTM) network. This allows the model to learn the sequential "narrative" of a user's history, identifying patterns that span multiple purchase events.
  3. Correlation-Sensitive Predictor: The final prediction layer integrates the LSTM's hidden state with the item correlation matrix. A hyperparameter $\alpha$ balances the influence of the sequential signal against the correlative associations, ensuring that the recommended items are both historically consistent and internally coherent.
The architecture of the Beacon model.

Experiments and Results

Beacon was evaluated against several state-of-the-art baselines, including DREAM and BSEQ, across three diverse datasets: TaFeng (grocery), Delicious (tags), and Foursquare (check-ins).

  • Accuracy and Utility: Beacon achieved superior performance across all metrics. On the TaFeng dataset, it reached an F1@5 of 6.36%, outperforming the next best baseline, DREAM (5.85%). Improvements in Half-life Utility (HLU) were even more pronounced, with gains of 10-20% over competing models.
  • Qualitative Impact: In the Delicious dataset, Beacon demonstrated an ability to suggest highly specific and relevant tag clusters. For a bookmark related to jQuery, Beacon recommended tags like web, design, programming, and javascript, whereas baseline models often suggested generic, less useful tags like digital or social.
  • Parameter Sensitivity: The study found that the optimal balance between sequential and correlative signals (determined by $\alpha$) typically falls between 0.2 and 0.6. Relying too heavily on either the sequence or the correlations alone leads to sub-optimal performance.

Conclusion

Beacon demonstrates that explicitly modeling item correlations within a sequential framework significantly enhances next-basket recommendations. By treating baskets as coherent coalitions rather than independent sets, the model aligns more closely with real-world user behavior. Future work may involve incorporating user-specific global embeddings and exploring dynamic correlation matrices that adapt to changing contexts or seasonal trends.

Summary updated

Abstract

Items adopted by a user over time are indicative of the underlying preferences. We are concerned with learning such preferences from observed sequences of adoptions for recommendation. As multiple items are commonly adopted concurrently, e.g., a basket of grocery items or a sitting of media consumption, we deal with a sequence of baskets as input, and seek to recommend the next basket. Intuitively, a basket tends to contain groups of related items that support particular needs. Instead of recommending items independently for the next basket, we hypothesize that incorporating information on pairwise correlations among items would help to arrive at more coherent basket recommendations. Towards this objective, we develop a hierarchical network architecture codenamed Beacon to model basket sequences. Each basket is encoded taking into account the relative importance of items and correlations among item pairs. This encoding is utilized to infer sequential associations along the basket sequence. Extensive experiments on three public real-life datasets showcase the effectiveness of our approach for the next-basket recommendation problem.
BibTeX
@inproceedings{le2019correlation,
  author = {Duc-Trong Le and Hady W. Lauw and Yuan Fang},
  title = {{Correlation-Sensitive Next-Basket Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/389},
}
Figure