Disentangled Contrastive Bundle Recommendation with Conditional Diffusion

Jiuqiang LiAAAI 2025 · DOI

TL;DR

This paper introduces DCBR, a novel framework for bundle recommendation that tackles noise and sparsity in user interaction data. By combining a conditional diffusion model for denoising user-bundle interactions with a dual-level disentangled contrastive learning approach, DCBR achieves state-of-the-art performance on benchmark datasets.

What Problem Does This Paper Address?

Bundle recommendation aims to suggest groups of items that users are likely to purchase together, enhancing user experience. However, existing methods face two main challenges:

  1. Noisy User Interactions: Real-world user behavior data is often noisy due to factors like accidental clicks or incomplete interactions, which can negatively impact recommendation quality.
  2. Data Sparsity: User interactions with bundles are typically very sparse, making it difficult for models to learn robust user preferences and capture complex relationships.

While self-supervised learning (SSL) has been used to mitigate sparsity, current approaches often struggle with the noise inherent in the data and can introduce further noise through the fusion of different data views (user-item, bundle-item, user-bundle).

Methods

The proposed Disentangled Contrastive Bundle Recommendation (DCBR) framework comprises three main components:

  1. Conditional Bundle Diffusion Model (CBDM): Inspired by diffusion models' success in data denoising, CBDM is adapted to denoise the user-bundle interaction graph. A key innovation is the Bundle Latent Consistency Constraint (BLCC), which ensures that the denoising process preserves genuine interaction information by maximizing the consistency between the original and denoised latent bundle representations. This helps mitigate the degradation of original interaction data.
  2. Triple-View Denoised Graph Learning: This module leverages graph neural networks (GNNs) to capture higher-order relationships across three views: user-bundle (UB), user-item (UI), and bundle-item (BI) interactions. It incorporates random noise perturbation to enhance robustness against noise in these interactions. The module adaptively fuses representations from these views using weighted average pooling.
  3. Dual-Level Disentangled Contrastive Learning (DDCL): To address the limitations of existing SSL methods that fuse representations before contrastive learning, DDCL operates on two levels:
  4. Inter-view level: Learns disentangled relationships between different views (UB, UI, BI) to capture global cross-view information.
  5. Intra-view level: Focuses on exploring fine-grained relationships within each individual view to enhance sensitivity to local interactions. By maximizing similarity between positive samples and minimizing it for negative samples across these disentangled views, DDCL generates robust contrastive signals that alleviate sparsity and noise issues.

The overall training objective combines the CBDM loss, the DDCL loss, and a Bayesian Personalized Ranking (BPR) loss for the recommendation task.

Datasets

The paper evaluates DCBR on three publicly available datasets:

  • MealRec: A dataset for meal recommendation.
  • MealRec+: An enhanced version of MealRec.
  • iFashion: A dataset for fashion outfit recommendation.

These datasets represent typical scenarios for bundle recommendation and vary in size and sparsity. The data partitioning follows standard practices from previous work.

Key Insights and Ideas

  • Diffusion for Denoising Interaction Graphs: Applying diffusion models not just for data generation but specifically for denoising noisy interaction graphs in recommendation is a novel approach. The conditional aspect and BLCC are crucial for preserving signal.
  • Disentangled Contrastive Learning is Key: The core idea that disentangling representations before contrastive learning, across both inter- and intra-view levels, is superior to fusing then contrasting is a significant contribution. This directly tackles the noise introduced by semantic gaps between views.
  • Triple-View Synergy: Effectively combining user-bundle, user-item, and bundle-item interactions through a denoised graph learning module provides a richer context than focusing on a single interaction type.
  • Robustness to Sparsity: The combination of denoising and disentangled contrastive learning demonstrably improves performance even in highly sparse scenarios, as shown in the robustness analysis.

Key Results

  • State-of-the-Art Performance: DCBR consistently outperforms several strong baselines (including BGCN, CrossCBR, and MultiCBR) across all three datasets and evaluation metrics (Recall@K and NDCG@K).
  • Significant Improvements: For example, on the iFashion dataset, DCBR shows improvements of up to 28.69% in Recall@10 and 31.19% in NDCG@10 compared to the best baseline.
  • Robustness: DCBR maintains its performance advantage across different levels of user interaction sparsity, outperforming baselines even on the sparsest user groups.
  • Ablation Studies: Removing individual components like BLCC, CBDM, or the inter/intra-view contrastive losses leads to significant performance degradation, validating the effectiveness of each proposed module.

Limitations and Future Work

  • Hyperparameter Sensitivity: While DCBR shows strong performance, the paper notes that hyperparameters like the BLCC loss weight ($\lambda_0$) and DDCL loss weight ($\lambda_1$) require careful tuning. Overly strong constraints can lead to performance degradation.
  • Computational Complexity: The triple-view graph learning and contrastive learning components add to the computational complexity, although the paper does not explicitly detail the trade-offs in terms of training time or inference speed compared to baselines.
  • Future Work: The paper implicitly suggests future work could explore more advanced diffusion model architectures or different contrastive learning strategies to further enhance disentanglement and robustness. Investigating the application of this framework to other recommendation tasks beyond bundles could also be a direction.

Summary updated

BibTeX
@inproceedings{li2025disentangled,
  author = {Jiuqiang Li},
  title = {{Disentangled Contrastive Bundle Recommendation with Conditional Diffusion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i11.33314},
}
Figure