MixRec: Individual and Collective Mixing Empowers Data Augmentation for Recommender Systems

Yi Zhang

Introduction

Modern recommender systems rely on learning high-quality embedding representations of users and items to predict future interactions. However, these systems frequently encounter the "data sparsity" problem: most users interact with only a fraction of available items, providing insufficient supervision signals for deep learning models. While Self-Supervised Learning (SSL) and Contrastive Learning (CL) have emerged as popular solutions to augment data, existing methods often rely on computationally expensive graph-based augmentations (like edge dropping) or manually intensive hyperparameter tuning.

To address these challenges, we present MixRec, a dual-mixing recommendation framework that performs data augmentation directly in the feature space. By utilizing individual and collective mixing strategies, MixRec provides a universal, efficient, and robust approach to alleviating data sparsity.

Methodology: The Dual Mixing Strategy

MixRec shifts the focus from graph-space augmentation to feature-space interpolation. This approach avoids the need for multiple passes through heavy Graph Neural Network (GNN) layers, ensuring linear time complexity and significantly reduced training overhead.

1. Individual Mixing (IM)

Individual Mixing generates a unique synthetic positive sample for a target user or item. By linearly interpolating a user’s embedding with a random user from the same training batch—while maintaining a high ratio for the original user—the model creates a "perturbed" version of the user that retains their core identity. This provides a new positive signal for the pair-wise recommendation loss.

2. Collective Mixing (CM)

Collective Mixing aims to capture group properties within a batch. Using a Dirichlet distribution, the framework combines all user embeddings in a batch into a single "collective" representation. This serves as a "hard negative" sample; it represents the general "wisdom of the crowd," which the model must learn to distinguish from a specific individual’s unique preferences.

3. Dual-Mixing Contrastive Learning (DMCL)

The DMCL loss function integrates these mixed samples to enhance representation consistency. It encourages the model to align the original user view with the individual mix (positive alignment) while maximizing the distance from the collective mix (negative discrimination).

Experimental Evaluation

The effectiveness of MixRec was evaluated on four diverse real-world datasets: Douban-Book, Yelp, Tmall, and Amazon-Book. These datasets range in sparsity from 99.72% to 99.94%, representing significant challenges for standard collaborative filtering.

Performance and Accuracy

MixRec consistently outperformed 20 state-of-the-art baselines, including LightGCN and SimGCL. Across all datasets, MixRec achieved improvements in Recall@20 ranging from 1.78% to 5.05%.

Efficiency and Scalability

One of MixRec’s primary advantages is its training speed. Because it operates in the feature space and avoids redundant graph encoding, it is significantly faster than traditional SSL methods. On the Tmall dataset, MixRec reached convergence in approximately 18–24 minutes, whereas some GCN-based competitors required over 13 hours to achieve similar results.

Sparsity Resistance

In analysis targeting "cold-start" or low-interaction users, MixRec demonstrated a 10-12% performance improvement over standard methods. This confirms that the dual-mixing strategy effectively fills the informational gaps in sparse data environments.

Discussion and Insights

  • Feature-Space Efficiency: Augmenting embeddings directly is more scalable than graph-based methods, which require re-encoding the entire graph multiple times.
  • Negative Sampling Innovation: Using the batch average (Collective Mixing) as a negative sample forces the model to learn sharper, more discriminative features by distinguishing individuals from the group average.
  • Usability: MixRec’s mixing parameters (such as the shape parameter $\alpha$) are remarkably stable across different datasets, reducing the need for exhaustive hyperparameter searches.

Conclusion

MixRec provides a streamlined yet powerful framework for data augmentation in recommender systems. By combining individual-level and group-level mixing, it achieves superior accuracy and sparsity resistance without the computational overhead of traditional graph-based augmentation. Future work may explore the application of these mixing strategies to more complex scenarios, such as social recommendation or knowledge-graph-enhanced systems.

Summary updated

Abstract

The core of the modern recommender systems lies in learning high-quality embedding representations of users and items to investigate their positional relations in the feature space. Unfortunately, data sparsity caused by difficult-to-access interaction data severely limits the effectiveness of recommender systems. Faced with such a dilemma, various types of self-supervised learning methods have been introduced into recommender systems in an attempt to alleviate the data sparsity through distribution modeling or data augmentation. However, most data augmentation relies on elaborate manual design, which is not only not universal, but the bloated and redundant augmentation process may significantly slow down model training progress. To tackle these limitations, we propose a novel Dual Mixing-based Recommendation Framework (MixRec) to empower data augmentation as we wish. Specifically, we propose individual mixing and collective mixing, respectively. The former aims to provide a new positive sample that is unique to the target (user or item) and to make the pair-wise recommendation loss benefit from it, while the latter aims to portray a new sample that contains group properties in a batch. The two mentioned mixing mechanisms allow for data augmentation with only one parameter that does not need to be set multiple times and can be done in linear time complexity. Besides, we propose the dual-mixing contrastive learning to maximize the utilization of these new-constructed samples to enhance the consistency between pairs of positive samples. Experimental results on four real-world datasets demonstrate the effectiveness of MixRec in terms of recommendation performance, training efficiency, sparsity resistance, and usability.
BibTeX
@inproceedings{zhang2025mixrec,
  author = {Yi Zhang and Yiwen Zhang},
  title = {{MixRec: Individual and Collective Mixing Empowers Data Augmentation for Recommender Systems}},
  booktitle = {The Web Conference},
  year = {2025},
  url = {https://openreview.net/forum?id=rNvMJEgcu2},
}
Figure