TraDiff: Network Traffic Matrix Prediction via Advanced Feature Fusion and Conditional Diffusion

Abstract

With the evolution of modern intelligent networks, accurate traffic matrix (TM) prediction has become critical for proactive network management, but remains challenging due to high-dimensional and non-stationary spatio-temporal dependencies. We propose TraDiff, an image-based conditional diffusion framework for TM prediction. TraDiff represents each TM as a grayscale image, extracts complementary spatial features using AlexNet, ResNet-50, and Vision Transformer (ViT), and integrates them through self-attention fusion. A temporal encoder then summarizes historical context, which is injected into a diffusion U-Net via feature-wise linear modulation (FiLM) for conditional generation. Experiments on the Abilene and GÉANT datasets demonstrate that TraDiff achieves the lowest MAE under all evaluated settings and remains competitive in RMSE across multiple forecasting horizons. In particular, on Abilene, TraDiff substantially reduces both MAE and RMSE compared with ViT-LSTM, with the largest relative improvements reaching 44.0% and 43.1%, respectively. On GEANT, it achieves the lowest MAE under all evaluated settings while maintaining competitive RMSE, further confirming its effectiveness.