GUESR: A Global Unsupervised Data-Enhancement with Bucket-Cluster Sampling for Sequential Recommendation

Yongqiang Han, Likang Wu, Hao Wang, Guifeng Wang, Mengdi Zhang, Zhi Li, Defu Lian, Enhong Chen

Introduction

Sequential Recommendation (SR) has become a cornerstone of modern digital platforms, powering personalized discovery on services ranging from streaming media to e-commerce. By modeling the chronological order of user interactions, SR systems attempt to predict the next item a user is likely to engage with. However, traditional SR models often suffer from "tunnel vision"—a limitation where the model focuses exclusively on a single user's history. This local focus makes models highly susceptible to data sparsity (items with few interactions) and data noise (accidental clicks).

To address these challenges, this paper introduces GUESR (Global Unsupervised Data-Enhancement with Bucket-Cluster Sampling). GUESR shifts the paradigm from local sequence modeling to a global perspective, leveraging item relationships across the entire dataset to provide a robust, denoised foundation for sequential recommendation.

The GUESR Framework

GUESR addresses the limitations of standard SR through a three-pronged architecture designed to capture global item associations, improve negative sampling, and model complex user interests.

1. Global Item Relationship Graph (GIRG)

Most contrastive learning methods in SR rely on simple data augmentations like masking or reordering items within a single sequence. GUESR instead constructs a Global Item Relationship Graph (GIRG). In this graph, items are nodes, and edges are weighted based on the frequency of their co-occurrence across all user sequences in the dataset.

To combat noise, GUESR employs a threshold-based pruning strategy. By deleting weak edges, the model naturally mitigates popularity bias, as popular items often accumulate many low-weight, coincidental associations. The resulting graph is then processed using Graph Contrastive Learning (GCL) with a LightGCN encoder. This process creates multiple augmented views of each item, forcing the model to learn representations that are consistent across different global contexts, effectively acting as a powerful regularizer.

2. Bucket-Cluster Sampling (BCS)

Contrastive learning performance is heavily dependent on the quality of negative samples. Standard models typically use random sampling, which often leads to "sampling bias"—where the model fails to distinguish between truly irrelevant items and those that are subtly different.

Bucket-Cluster Sampling (BCS) addresses this by grouping items into "buckets" using a combination of coarse-grained attributes and K-means clustering on their current embeddings. By sampling negatives from diverse buckets, GUESR ensures the model is exposed to a representative set of alternatives, forcing it to learn more nuanced boundaries between item categories.

3. Multi-Interest Extraction

Users rarely have a single, linear interest. A single user might be browsing for both "sci-fi novels" and "cycling gear" simultaneously. GUESR utilizes a Capsule Network (CapsNet) with dynamic routing to extract multiple interest vectors for each user.

To make a final prediction, the model employs a target-attention mechanism. This mechanism weights the extracted interest vectors based on the specific candidate item being evaluated. For example, if the candidate item is a bike helmet, the model will prioritize the user's "cycling" interest vector over their "reading" interest vector.

Experimental Evaluation

The authors evaluated GUESR against several state-of-the-art baselines, including CL4SRec and S3Rec, across four diverse datasets: ML-1M (movies), Amazon Sports, Yelp (business reviews), and Amazon Books.

Key Findings

  • Superior Performance: GUESR consistently outperformed all baselines. On the ML-1M dataset, GUESR achieved a Recall@10 of 0.2157, a significant improvement over the second-best model, CL4SRec (0.2034).
  • The Power of Global Signals: Ablation studies revealed that the Global Contrastive Learning (GCL) component is the most critical element of the framework. Removing GCL resulted in a performance drop of up to 29% on the Amazon Books dataset.
  • Plug-and-Play Versatility: One of the most significant findings is that GUESR’s data-enhancement module can be integrated into existing models like SASRec or GRU4Rec, providing a consistent performance boost across different architectures.

Conclusion and Future Work

GUESR demonstrates that incorporating global item relationships can significantly improve the robustness and accuracy of sequential recommendation systems. By pruning noisy edges in a global graph and utilizing sophisticated bucketed sampling, the model overcomes the inherent limitations of sparse, noisy user data.

While highly effective, the authors note that scaling the global item graph to platforms with millions of items remains a computational challenge. Future research may focus on making the GIRG dynamic to account for seasonal trends and evolving item relationships, as well as optimizing hyperparameter tuning for the noise threshold ($\epsilon$) and sampling depth ($D$).

Summary updated

BibTeX
@inproceedings{han2023guesr,
  author = {Yongqiang Han and Likang Wu and Hao Wang and Guifeng Wang and Mengdi Zhang and Zhi Li and Defu Lian and Enhong Chen},
  title = {{GUESR: A Global Unsupervised Data-Enhancement with Bucket-Cluster Sampling for Sequential Recommendation}},
  booktitle = {DASFAA},
  year = {2023},
  url = {https://openreview.net/forum?id=A1ptQr5lc2},
}
Figure