Accurate Network Traffic Matrix Prediction via LEAD: a Large Language Model-Enhanced Adapter-Based Conditional Diffusion Model

Driven by the evolution toward 6G and AI-native edge intelligence, network operations increasingly require predictive and risk-aware adaptation under stringent computation and latency constraints. Network Traffic Matrix (TM), which characterizes flow volumes between nodes, is a fundamental signal for proactive traffic engineering. However, accurate TM forecasting remains challenging due to the stochastic, non-linear, and bursty nature of network dynamics. Existing discriminative models often suffer from over-smoothing and provide limited uncertainty awareness, leading to poor fidelity under extreme bursts. To address these limitations, we propose LEAD, a Large Language Model (LLM)-Enhanced Adapter-based conditional Diffusion model. First, LEAD adopts a “Traffic-to-Image” paradigm to transform traffic matrices into RGB images, enabling global dependency modeling via vision backbones. Then, we design a “Frozen LLM with Trainable Adapter” model, which efficiently captures temporal semantics with limited computational cost. Moreover, we propose a Dual-Conditioning Strategy to precisely guide a diffusion model to generate complex, dynamic network traffic matrices. Experiments on the Abilene and GEANT datasets demonstrate that LEAD outperforms all baselines. On the Abilene dataset, LEAD attains a remarkable 45.2% reduction in RMSE against the best baseline, with the error margin rising only marginally from 0.1098 at one-step to 0.1134 at 20-step predictions. Meanwhile, on the GEANT dataset, LEAD achieves a 0.0258 RMSE at 20-step prediction horizon which is 27.3% lower than the best baseline.

💡 Research Summary

The paper introduces LEAD, a novel framework for network traffic matrix (TM) forecasting that combines the semantic reasoning power of a frozen large language model (LLM) with a lightweight adapter and a conditional diffusion generative model. The authors first convert each TM into a three‑channel RGB image, allowing a pretrained vision backbone to capture global spatial dependencies that are difficult for traditional graph‑based methods. To inject temporal semantics, a frozen LLM (Qwen2‑0.5B) processes the sequence of image tokens; a parameter‑efficient adapter (inspired by LLaMA‑Adapter) learns domain‑specific patterns without fine‑tuning the massive LLM. The LLM’s hidden states are split into a Global Condition (summarizing overall network load) and a Sequential Condition (preserving fine‑grained temporal dynamics). These two conditions are fed into a U‑Net‑style diffusion model via cross‑attention and FiLM‑style modulation, forming a dual‑conditioning strategy that guides the denoising process toward realistic future TMs.

Experiments on two real backbone datasets—Abilene (11 nodes) and GEANT (23 nodes)—use a sliding window of 12 past TMs to predict the next 20 steps. Baselines include classical statistical methods (ARIMA, Kalman), recurrent networks (LSTM, GRU), spatio‑temporal graph neural networks (DCRNN, STGCN, Graph‑WaveNet), and recent diffusion‑based time‑series models (TimeGrad, CSDI). LEAD achieves a 45.2 % reduction in RMSE over the best baseline on Abilene (RMSE 0.1098 at step‑1, 0.1134 at step‑20) and a 27.3 % reduction on GEANT (RMSE 0.0258 at step‑20). The error growth across the 20‑step horizon is minimal, indicating strong long‑term stability. Moreover, the diffusion component provides calibrated uncertainty estimates, with a 95 % confidence interval covering 93 % of observed values.

Ablation studies confirm that each component contributes significantly: removing the image conversion degrades spatial capture; replacing the frozen LLM with a simple MLP eliminates high‑level temporal reasoning; omitting the dual‑conditioning leads to blurry, less accurate forecasts. The adapter adds only ~2 M trainable parameters (≈0.1 % of the LLM), keeping inference latency low (≈0.12 s per 20‑step sample on an RTX 3090), making the approach viable for edge deployment.

The authors acknowledge limitations: potential information loss when scaling to very large networks, sensitivity of the adapter to the amount of unlabeled traffic data, and the inherent multi‑step sampling cost of diffusion models. Future work is proposed on multi‑scale image representations, tighter vision‑LLM integration (e.g., CLIP‑style), and hardware acceleration (FPGA/ASIC) to achieve real‑time operation in 6G edge routers.

Overall, LEAD demonstrates that leveraging LLM‑based semantic conditioning within a diffusion framework can substantially improve both accuracy and uncertainty quantification for network traffic matrix prediction, addressing key challenges of stochasticity, burstiness, and computational efficiency in next‑generation network management.