InfoDCL: Informative Noise Enhanced Diffusion Based Contrastive Learning

Xufeng Liang, Zhida Qin, Chong Zhang, Tianyu Huang, Gangyi DingarXiv 2025 · arXiv · OpenReview

Abstract

Contrastive learning has demonstrated promising potential in recommender systems. Existing methods typically construct sparser views by randomly perturbing the original interaction graph, as they have no idea about the authentic user preferences. Owing to the sparse nature of recommendation data, this paradigm can only capture insufficient semantic information. To address the issue, we propose InfoDCL, a novel diffusion-based contrastive learning framework for recommendation. Rather than injecting randomly sampled Gaussian noise, we employ a single-step diffusion process that integrates noise with auxiliary semantic information to generate signals and feed them to the standard diffusion process to generate authentic user preferences as contrastive views. Besides, based on a comprehensive analysis of the mutual influence between generation and preference learning in InfoDCL, we build a collaborative training objective strategy to transform the interference between them into mutual collaboration. Additionally, we employ multiple GCN layers only during inference stage to incorporate higher-order co-occurrence information while maintaining training efficiency. Extensive experiments on five real-world datasets demonstrate that InfoDCL significantly outperforms state-of-the-art methods. Our InfoDCL offers an effective solution for enhancing recommendation performance and suggests a novel paradigm for applying diffusion method in contrastive learning frameworks.

Introduction

Modern recommendation engines frequently leverage Contrastive Learning (CL) to handle the inherent sparsity of user-item interaction data. The standard CL paradigm involves creating multiple augmented "views" of a user's profile—typically through random dropout or Gaussian noise—and training the model to maximize agreement between these views. However, this "blind" perturbation often fails because it ignores the underlying distribution of authentic user preferences. In sparse datasets, random noise can easily destroy the subtle semantic signals necessary for accurate recommendation. InfoDCL addresses these limitations by replacing random perturbations with a generative process that creates "informative noise."

Methodology: The InfoDCL Framework

The InfoDCL framework consists of several key components designed to transform noise into meaningful training signals:

1. Preference Signal Network (PSNet)

The core innovation is the PSNet, which generates informative noise by integrating auxiliary metadata (such as item-item similarity) with raw noise. Using Spectral Rectification via Singular Value Decomposition (SVD), PSNet approximates a single-step diffusion process. This ensures that the injected noise is not merely random but carries semantic weight, acting as a "semantic gradient" that guides the model toward authentic user preferences.

2. Informative Diffusion Process

The refined noise from PSNet is fed into a diffusion model's forward process. During the denoising phase, the model generates contrastive views that are "preference-aware." Unlike traditional CL, which breaks data to learn, InfoDCL generates high-quality synthetic examples that reflect likely user interests.

Overall architecture of the InfoDCL framework.

3. Collaborative Training and Inference

To resolve potential interference between the generative (diffusion) and discriminative (ranking) tasks, InfoDCL employs a Collaboration Balance Loss. This synchronizes the reconstruction loss of the diffusion process with the Bayesian Personalized Ranking (BPR) loss. Furthermore, to maintain efficiency, Graph Convolutional Network (GCN) layers are utilized exclusively during the inference stage, allowing the model to capture higher-order relationships without slowing down training.

Key Theoretical Insights

  • SVD as a Diffusion Shortcut: The authors utilize the Davis-Kahan theorem to justify using SVD as a way to approximate a single-step diffusion process. This allows for the efficient generation of "informative noise" without the computational overhead of a full multi-step generative chain.
  • The Semantic Gradient: The framework mathematically ensures that informative noise acts as a gradient, guiding embeddings toward higher expected preference scores compared to standard Gaussian noise.

Experimental Evaluation

The model was tested on five real-world datasets: ML-1M, Amazon-Office, Amazon-Baby, Amazon-Electronics, and Taobao. These datasets represent various scales and levels of sparsity.

Performance Gains

InfoDCL consistently outperformed state-of-the-art baselines like LightGCN, SimGCL, and DiffRec. Notable results include:

  • ML-1M: A 42.7% improvement in Recall@20 over the best baseline.
  • Taobao: A 24.2% improvement in Recall@20, demonstrating robustness in extremely sparse environments.

Ablation Study

Ablation experiments confirm that each component of InfoDCL—Spectral Rectification (SR), Contextual Re-encoding (CR), PSNet, and Collaboration Balance Loss (CBL)—contributes significantly to the final performance.

Ablation study results across multiple datasets.

Conclusion

InfoDCL represents a shift from perturbative to generative contrastive learning in recommender systems. By making noise "meaningful" through a diffusion-based approach, it effectively captures semantic information even in sparse data scenarios. Future work may explore integrating multimodal metadata, such as product images or text, to further enrich the informative noise generation process.

Summary updated

Abstract

Contrastive learning has demonstrated promising potential in recommender systems. Existing methods typically construct sparser views by randomly perturbing the original interaction graph, as they have no idea about the authentic user preferences. Owing to the sparse nature of recommendation data, this paradigm can only capture insufficient semantic information. To address the issue, we propose InfoDCL, a novel diffusion-based contrastive learning framework for recommendation. Rather than injecting randomly sampled Gaussian noise, we employ a single-step diffusion process that integrates noise with auxiliary semantic information to generate signals and feed them to the standard diffusion process to generate authentic user preferences as contrastive views. Besides, based on a comprehensive analysis of the mutual influence between generation and preference learning in InfoDCL, we build a collaborative training objective strategy to transform the interference between them into mutual collaboration. Additionally, we employ multiple GCN layers only during inference stage to incorporate higher-order co-occurrence information while maintaining training efficiency. Extensive experiments on five real-world datasets demonstrate that InfoDCL significantly outperforms state-of-the-art methods. Our InfoDCL offers an effective solution for enhancing recommendation performance and suggests a novel paradigm for applying diffusion method in contrastive learning frameworks.
BibTeX
@misc{liang2025infodcl,
  author = {Xufeng Liang and Zhida Qin and Chong Zhang and Tianyu Huang and Gangyi Ding},
  title = {{InfoDCL: Informative Noise Enhanced Diffusion Based Contrastive Learning}},
  year = {2025},
  eprint = {2512.16576},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.16576},
}
Figure