A Survey on Bundle Recommendation: Methods, Applications, and Challenges

In recent years, bundle recommendation systems have gained significant attention in both academia and industry due to their ability to enhance user experience and increase sales by recommending a set of items as a bundle rather than individual items. This survey provides a comprehensive review on bundle recommendation, beginning by a taxonomy for exploring product bundling. We classify it into two categories based on bundling strategy from various application domains, i.e., discriminative and generative bundle recommendation. Then we formulate the corresponding tasks of the two categories and systematically review their methods: 1) representation learning from bundle and item levels and interaction modeling for discriminative bundle recommendation; 2) representation learning from item level and bundle generation for generative bundle recommendation. Subsequently, we survey the resources of bundle recommendation including datasets and evaluation metrics, and conduct reproducibility experiments on mainstream models. Lastly, we discuss the main challenges and highlight the promising future directions in the field of bundle recommendation, aiming to serve as a useful resource for researchers and practitioners. Our code and datasets are publicly available at https://github.com/WUT-IDEA/bundle-recommendation-survey.

💡 Research Summary

This survey provides a comprehensive overview of the rapidly evolving field of bundle recommendation, a paradigm that extends traditional item recommendation by suggesting a set of items (a bundle) to a single user. The authors begin by positioning bundle recommendation within the broader landscape of recommender systems, distinguishing it from item‑to‑item, group‑to‑item, and complex set recommendation tasks. They introduce a clear taxonomy that separates bundle recommendation into two major categories based on the underlying strategy: discriminative bundle recommendation, which matches existing bundles to users, and generative bundle recommendation, which creates new bundles on the fly.

For discriminative bundle recommendation, the survey systematically reviews the literature on representation learning at both the bundle level and the item level, highlighting how modern deep learning encoders such as Graph Neural Networks (GNNs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs) are employed to capture user‑bundle (U‑B) and bundle‑item (B‑I) interactions. The authors discuss interaction modeling techniques, including cross‑attention mechanisms, contrastive learning, and knowledge distillation, and they emphasize probabilistic approaches like Determinantal Point Processes (DPP) that explicitly model diversity and redundancy among items within a bundle.

In the generative branch, the paper surveys methods that generate novel bundles from item representations. It covers item‑level embedding techniques ranging from Latent Dirichlet Allocation (LDA) to large language models (LLMs), and it examines sequence‑to‑sequence, Transformer, and Generative Adversarial Network (GAN) architectures for bundle construction. The authors describe how reinforcement learning, diversity‑aware sampling, and multi‑modal fusion (text, image, price) are integrated to satisfy business constraints such as bundle size, category balance, and price discount.

The survey also compiles an extensive list of publicly available datasets across five domains—e‑commerce, fashion, entertainment, food, and travel—detailing bundle characteristics (size distribution, category diversity, pricing schemes) and providing a taxonomy of evaluation metrics. Beyond traditional accuracy‑oriented measures (Precision@k, Recall@k, NDCG), the authors incorporate bundle‑level coverage, diversity, revenue impact, and user satisfaction, arguing for a multi‑objective evaluation framework.

To assess reproducibility, the authors conduct experiments on several state‑of‑the‑art discriminative models (e.g., Bi‑GNN, Bundle‑CF) and generative models (e.g., Seq2Seq‑GAN, Transformer‑Bundle) under uniform data splits and metric suites. Results show that graph‑based dual‑representation models achieve the highest predictive accuracy for existing‑bundle matching, while Transformer‑based generators excel at producing diverse, novel bundles at the cost of higher computational overhead.

Four major challenges are identified: (1) modeling complex intra‑bundle item interactions, (2) data sparsity and limited labeled bundles in many domains, (3) the lack of a unified multi‑objective evaluation protocol that balances accuracy, diversity, revenue, and fairness, and (4) scalability and online learning constraints for real‑time deployment. The authors propose future research directions, including (i) integrating hypergraph and large‑scale pre‑trained models for richer interaction modeling, (ii) multi‑task and multi‑modal learning to improve cross‑domain generalization, (iii) reinforcement learning or Bayesian optimization for bundle generation policy refinement, (iv) fairness‑aware and privacy‑preserving bundle recommendation, and (v) automated A/B testing pipelines with adaptive metric selection.

In summary, this survey not only consolidates the latest methodological advances, datasets, and evaluation practices for both discriminative and generative bundle recommendation but also outlines a clear roadmap for addressing current limitations. It serves as a valuable reference for researchers seeking to push the boundaries of bundle recommendation and for practitioners aiming to deploy effective, scalable bundle recommendation systems in real‑world applications.