Enhancing Sequential Recommendation with Global Diffusion

Mingxuan Luo, Yang Li, Chen LinAAAI 2025 · DOI

TL;DR

This paper introduces GlobalDiff, a novel plug-and-play framework that enhances sequential recommendation models by integrating a diffusion model. GlobalDiff leverages the global, non-sequential structure of item interactions to compensate for inconsistencies in local sequences, leading to significant improvements in next-item prediction accuracy.

What Problem Does This Paper Address?

Traditional sequential recommendation (SR) models excel at capturing temporal dependencies in user interaction sequences. However, they often struggle with "non-sequential item selection problems," where a user's choice might be influenced by external factors (like promotions or social media trends) rather than purely personal interest. This can lead to recommendations that are locally coherent but globally suboptimal. For instance, a user interested in horror movies might pick a popular action movie, disrupting the expected sequential pattern. Existing SR models, which primarily rely on local sequential context, can be misled by such inconsistencies.

Methods

GlobalDiff proposes a framework that stacks a diffusion model on top of an existing sequential recommendation model. The core idea is to use the diffusion model to capture and restore the "global non-sequential data structure" of the item universe, thereby balancing personal interest with broader item popularity and correlations.

Key novel techniques include:

  • Training Sample Construction: To adapt diffusion models (which typically work with continuous data) to discrete items, GlobalDiff transforms masked items into vectorized samples. These samples encode both global item associations and positional information within the sequence, creating a richer training set.
  • Guided Reverse Approximator: The diffusion model's reverse (denoising) process is guided by the contextual information from the pre-trained sequential model. This ensures that the diffusion model learns to reconstruct personalized item representations, improving training stability and mitigating issues like mode collapse.
  • Inference Ensemble: At inference time, GlobalDiff combines the predictions from the sequential model (representing personal interest) and the diffusion model (representing global structure) using a voting mechanism. This approach balances local and global information and reduces the risk of over-fitting to unusual item selections.

The framework is designed to be "plug-and-play," meaning it can be integrated with various pre-trained sequential models without altering their original architecture or training objectives.

Datasets

The authors evaluated GlobalDiff on three publicly available datasets:

  • MovieLens 1M (ML-1M): Movie ratings data.
  • Amazon-Beauty: User interactions with beauty products from an e-commerce platform.
  • KuaiRec: Interactions from a short-video platform.

These datasets cover different domains and exhibit varying levels of sparsity and sequence lengths, allowing for a comprehensive evaluation.

Key Insights and Ideas

  • Global Context is Crucial: The paper strongly argues that user item selection is influenced by both personal interest (captured by sequential models) and global item structure (popularity, correlations). Ignoring the latter leads to suboptimal recommendations.
  • Diffusion Models as Complementary Tools: Diffusion models, when integrated thoughtfully, can effectively capture and leverage this global item structure, acting as a powerful complement to sequential models.
  • "Plug-and-Play" is Achievable: The proposed framework demonstrates that advanced diffusion techniques can be seamlessly integrated with existing SR architectures without requiring significant modifications, making it practical for researchers and practitioners.
  • Balancing Local and Global: The inference ensemble method is key to effectively combining the strengths of both sequential and diffusion models, leading to more robust predictions.

Key Results

  • Next-Item Prediction: GlobalDiff consistently improved the performance of backbone sequential models across all datasets, achieving an average improvement of 9.67% in metrics like HR@K and NDCG@K.
  • State-of-the-Art Enhancement: GlobalDiff enhanced the SOTA CBiT model by an average of 5.17%.
  • Long-Term Prediction: The framework also significantly improved long-term prediction accuracy (predicting items further into the future), showing an average improvement of 9.40% in HR@20 across three backbone models.
  • Ablation Studies:
    • Guided representation in the reverse approximator was crucial, leading to substantial performance gains compared to using only target representation or no guidance.
  • The inference ensemble (voting) proved effective, even with limited diffusion model training, demonstrating its ability to correct sequential model errors.
  • The impact of diffusion steps (T) on performance was found to be insignificant, suggesting that efficient inference is possible.

Limitations and Future Work

  • Time Complexity: The paper acknowledges that the initial aggregation step for computing item-item similarity has a time complexity of O(M²), where M is the number of items. While they propose a cluster-item similarity approach as a scalable alternative, the exact pair-wise computation can be a bottleneck for very large item sets.
  • Future Work: The authors plan to explore more ways to reduce the complexity of GlobalDiff, indicating that efficiency remains an area for further research.

Summary updated

BibTeX
@inproceedings{luo2025enhancing,
  author = {Mingxuan Luo and Yang Li and Chen Lin},
  title = {{Enhancing Sequential Recommendation with Global Diffusion}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i12.33341},
}
Figure