DG-STMTL: A Novel Graph Convolutional Network for Multi-Task Spatio-Temporal Traffic Forecasting

TL;DR

DG-STMTL addresses the challenge of multi-task spatio-temporal traffic forecasting by integrating a hybrid adjacency generation mechanism that blends stable prior structure with learnable dynamics, guided by a task-specific gating scheme. The Cross-Task Knowledge Exchange unit enables dynamic, graph-based information sharing across tasks, while a two-stage group-wise spatio-temporal convolution captures both short- and long-term dependencies. Across three real-world datasets, DG-STMTL delivers state-of-the-art results and robust ablation performance, highlighting the value of combining static and dynamic graphs with structured cross-task interactions. The framework has broad practical implications for intelligent transportation and other domains requiring accurate, scalable spatio-temporal multi-task predictions, and opens avenues for future explainability and extension to diverse STMTL problems.

Abstract

Spatio-temporal traffic prediction is crucial in intelligent transportation systems. The key challenge of accurate prediction is how to model the complex spatio-temporal dependencies and adapt to the inherent dynamics in data. Traditional Graph Convolutional Networks (GCNs) often struggle with static adjacency matrices that introduce domain bias or learnable matrices that may be overfitting to specific patterns. This challenge becomes more complex when considering Multi-Task Learning (MTL). While MTL has the potential to enhance prediction accuracy through task synergies, it can also face significant hurdles due to task interference. To overcome these challenges, this study introduces a novel MTL framework, Dynamic Group-wise Spatio-Temporal Multi-Task Learning (DG-STMTL). DG-STMTL proposes a hybrid adjacency matrix generation module that combines static matrices with dynamic ones through a task-specific gating mechanism. We also introduce a group-wise GCN module to enhance the modelling capability of spatio-temporal dependencies. We conduct extensive experiments on two real-world datasets to evaluate our method. Results show that our method outperforms other state-of-the-arts, indicating its effectiveness and robustness.

Figures (7)

• Figure 1: Overview of the Dynamic Graph-based Multi-Task Learning Framework Architecture
• Figure 2: Illustration of several potential combinations of the prior adjacency matrix ($A_{p1}$, $A_{p2}$, $A_{p3}$, $A_{p4}$). Each configuration integrates spatial ($A_S$), temporal ($A_T$), and spatio-temporal ($A_{ST}$) components across three time steps ($t_1$, $t_2$, $t_3$). For instance, $A_{p1}$ focuses on temporal transitions ($A_T$) between specific pairs of consecutive time steps (e.g., $t_1$-$t_2$ and $t_2$-$t_1$), while spatio-temporal relationships ($A_{ST}$) are emphasized for other consecutive pairs (e.g., $t_2$-$t_3$ and $t_3$-$t_2$). By comparison, other configurations, such as $A_{p4}$, provide a more uniform emphasis on spatio-temporal relationships across all time steps.
• Figure 3: Illustration of the Group-wise Spatio-Temporal Graph Convolution Module
• Figure 4: Heatmaps of average traffic flow and speed across all nodes for PEMSD4 and PEMSD8 over two days
• Figure 5: Comparison of Training and Validation Loss Across Variant1, Variant2, Variant3 and Proposed model