U-STS-LLM unifies cellular traffic forecasting and imputation in single LLM
Researchers propose a spatio-temporally steered language model that handles both traffic prediction and missing-data imputation for cellular networks, outperforming specialized graph neural networks while training faster.
A new preprint describes U-STS-LLM, a framework that adapts large language models to predict cellular network traffic and fill in missing sensor data simultaneously. The paper, posted to arXiv on May 13, addresses two historically separate problems — forecasting future network load and imputing gaps from sensor failures — using a single unified architecture built on a pre-trained LLM backbone.
The core technical contribution is a Dynamic Spatio-Temporal Attention Bias Generator. It synthesizes what the authors call a "persistent functional graph" with transient node states, explicitly steering the LLM's attention mechanism to respect the spatial structure of cell towers and the temporal evolution of traffic. The model uses Low-Rank Adaptation (LoRA) to fine-tune a partially frozen LLM backbone, paired with a Gated Adaptive Fusion layer that blends structural guidance with learned representations. Training happens under a multi-task objective that jointly optimizes for both forecasting and imputation.
What stands out
- 01State-of-the-art on both tasks. Experiments on real-world cellular datasets show U-STS-LLM outperforms existing Spatio-Temporal Graph Neural Networks (STGNNs) in long-horizon forecasting and high-missing-rate imputation — the first framework to claim top performance on both simultaneously.
- 02Faster training, stable convergence. The LoRA-tuned backbone trains more efficiently than full-parameter STGNNs, and the explicit attention bias prevents the unstable convergence that plagues naive LLM adaptations to structured data.
- 03Unified representation learning. By training on forecasting and imputation together, the model learns a holistic spatio-temporal representation that transfers between tasks, reducing the need for task-specific architectures.
- 04Blueprint for non-linguistic domains. The authors position U-STS-LLM as a template for applying foundation models to structured, non-text data — cellular traffic, sensor networks, transportation grids — where spatial and temporal dependencies matter.