Unveiling the Power of Noise Priors: Enhancing Diffusion Models for Mobile Traffic Prediction

Zhi Sheng, Daisy Yuan, Jingtao Ding, Qi Yan, Xi Zheng, Yue Sun, Yong Li

Introduction

Predicting mobile network traffic is a critical challenge for modern urban planning and network optimization. The primary difficulty lies in the "non-stationary" nature of cellular data, which intertwines regular human activity patterns (e.g., daily commutes) with stochastic, unpredictable spikes caused by public events or emergencies.

While diffusion models have emerged as a powerful solution for capturing complex temporal dynamics due to their probabilistic framework, they typically rely on a standard assumption: that noise is pure, independent Gaussian randomness. This paper argues that for structured temporal data like mobile traffic, this "one-size-fits-all" noise assumption is sub-optimal. It forces the denoising network to reconstruct all structural information from scratch, ignoring the inherent rhythms already present in the data.

The NPDiff Framework

To address these limitations, the authors propose NPDiff, a novel framework that shifts the focus from architectural complexity to the fundamental role of noise. Instead of starting the diffusion process with pure white noise, NPDiff utilizes a "noise prior" derived from the underlying data dynamics.

Comparison of existing diffusion frameworks vs. the proposed Noise Prior framework.

As illustrated in the comparison above, the core innovation is the Noise Prior Estimation module. This module decomposes the target noise into a predictable prior and a learned residual, allowing the model to leverage known structural patterns as an informed starting point.

Methodology: Noise Prior Estimation

The NPDiff architecture redefines the iterative denoising step by fusing two distinct components:

  1. Noise Prior ($\tilde{\epsilon}$): This component is calculated directly from data dynamics. It captures periodic dynamics (using Fast Fourier Transform to identify long-term cycles) and local dynamics (using lag-1 values to capture immediate dependencies).
  2. Residual Noise: This represents the unpredictable variations that the denoising network ($\epsilon_\theta$) is specifically trained to estimate.
Detailed architecture of the NPDiff iterative denoising process, showing the fusion of the noise prior and the learned residual.

These components are combined using an adjustable hyperparameter $\lambda$. By providing the model with a prior that already captures a significant portion of the data's similarity, the diffusion process becomes more efficient, allowing the neural network to focus its capacity on learning the most complex, irregular variations.

Experimental Evaluation

The researchers evaluated NPDiff across four large-scale cellular network traffic datasets from Beijing, Nanjing, and Shanghai. The framework demonstrated significant improvements across several key metrics:

  • Accuracy: NPDiff outperformed 13 state-of-the-art baselines. Compared to standard diffusion models like CSDI, NPDiff reduced Mean Absolute Error (MAE) by an average of 41.6% and Root Mean Square Error (RMSE) by 54.4%.
  • Efficiency: Because the model does not need to relearn basic periodic patterns, it achieves faster training convergence and reaches optimal solutions in fewer iterations.
  • Robustness: The model exhibited higher stability under random noise perturbations and produced predictions with significantly narrower confidence intervals, enhancing its reliability for real-world network management.

Conclusion and Future Work

NPDiff demonstrates that noise in temporal diffusion models is not merely random but carries structural information shaped by data dynamics. By incorporating this insight through a noise prior, NPDiff achieves a "plug-and-play" enhancement that is compatible with various denoising backbones.

Future research directions include developing adaptive fusion mechanisms to dynamically adjust the influence of the noise prior across different time steps and exploring more sophisticated priors that incorporate external factors like weather or social events to further improve prediction accuracy in highly volatile scenarios.

Summary updated

Abstract

Accurate prediction of mobile traffic,i.e., network traffic from cellular base stations, is crucial for optimizing network performance and supporting urban development. However, the non-stationary nature of mobile traffic, driven by human activity and environmental changes, leads to both regular patterns and abrupt variations. Diffusion models excel in capturing such complex temporal dynamics due to their ability to capture the inherent uncertainties. Most existing approaches prioritize designing novel denoising networks but often neglect the critical role of noise itself, potentially leading to sub-optimal performance. In this paper, we introduce a novel perspective by emphasizing the role of noise in the denoising process. Our analysis reveals that noise fundamentally shapes mobile traffic predictions, exhibiting distinct and consistent patterns. We propose NPDiff, a framework that decomposes noise into prior and residual components, with the prior derived from data dynamics, enhancing the model's ability to capture both regular and abrupt variations. NPDiff can seamlessly integrate with various diffusion-based prediction models, delivering predictions that are effective, efficient, and robust. Extensive experiments demonstrate that it achieves superior performance with an improvement over 30%, offering a new perspective on leveraging diffusion models in this domain. We provide code and data at https://github.com/tsinghua-fib-lab/NPDiff.
BibTeX
@inproceedings{sheng2025unveiling,
  author = {Zhi Sheng and Daisy Yuan and Jingtao Ding and Qi Yan and Xi Zheng and Yue Sun and Yong Li},
  title = {{Unveiling the Power of Noise Priors: Enhancing Diffusion Models for Mobile Traffic Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=L9X5QrOMaI},
}
Figure