UniST: A Pre-Trained Unified Spatio-Temporal Multi-Task Framework for Network Traffic Analysis

Abstract

Spatio-temporal traffic data analysis is essential for 5G and intelligent network systems. However, existing methods often employ isolated models, neglecting shared spatiotemporal correlations and incurring prohibitive computational overhead. To address these issues, we propose UniST, a unified spatio-temporal multi-task learning framework for simultaneous prediction, imputation, and anomaly detection. The framework features a dual-pathway mechanism: a Spatio-Temporal Graph Embedding path explicitly models data with physical topology, while a Frequency Domain Enhancement path captures periodic variations in graph-free scenarios. These extracted features are processed by a shared GPT-2-based Transformer backbone to capture evolving temporal patterns. To mitigate gradient conflicts during joint multi-task optimization, we incorporate the Projecting Conflicting Gradients (PCGrad) algorithm to promote balanced convergence. Extensive experiments on four real-world datasets show that UniST achieves competitive and balanced performance across multiple tasks while reducing the overall training cost compared with deploying isolated task-specific models. Furthermore, few-shot experiments confirm its robust cross-domain transferability.