Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SpoT-Mamba: Learning Long-Range Dependency on Spatio-Temporal Graphs with Selective State Spaces

Jinhyeok Choi, Heehyeon Kim, Minhyeong An, Joyce Jiyoung Whang

TL;DR

SpoT-Mamba tackles the challenge of long-range dependency modeling in spatio-temporal graph forecasting by integrating a Mamba-based sequence model with node-specific walk embeddings. It constructs robust node representations from multi-way walk sequences and applies temporal scans via Mamba blocks, augmented by transformer-based global context, to forecast future node attributes efficiently. The approach delivers state-of-the-art performance on the PEMS04 traffic dataset, showing strong improvements in MAPE and competitive gains across MAE and RMSE while maintaining linear scaling in sequence length. This work offers a scalable, selective-state-space alternative to attention-heavy STGNNs, with practical implications for long-horizon forecasting in dynamic graph-structured systems.

Abstract

Spatio-temporal graph (STG) forecasting is a critical task with extensive applications in the real world, including traffic and weather forecasting. Although several recent methods have been proposed to model complex dynamics in STGs, addressing long-range spatio-temporal dependencies remains a significant challenge, leading to limited performance gains. Inspired by a recently proposed state space model named Mamba, which has shown remarkable capability of capturing long-range dependency, we propose a new STG forecasting framework named SpoT-Mamba. SpoT-Mamba generates node embeddings by scanning various node-specific walk sequences. Based on the node embeddings, it conducts temporal scans to capture long-range spatio-temporal dependencies. Experimental results on the real-world traffic forecasting dataset demonstrate the effectiveness of SpoT-Mamba.

SpoT-Mamba: Learning Long-Range Dependency on Spatio-Temporal Graphs with Selective State Spaces

TL;DR

SpoT-Mamba tackles the challenge of long-range dependency modeling in spatio-temporal graph forecasting by integrating a Mamba-based sequence model with node-specific walk embeddings. It constructs robust node representations from multi-way walk sequences and applies temporal scans via Mamba blocks, augmented by transformer-based global context, to forecast future node attributes efficiently. The approach delivers state-of-the-art performance on the PEMS04 traffic dataset, showing strong improvements in MAPE and competitive gains across MAE and RMSE while maintaining linear scaling in sequence length. This work offers a scalable, selective-state-space alternative to attention-heavy STGNNs, with practical implications for long-horizon forecasting in dynamic graph-structured systems.

Abstract

Spatio-temporal graph (STG) forecasting is a critical task with extensive applications in the real world, including traffic and weather forecasting. Although several recent methods have been proposed to model complex dynamics in STGs, addressing long-range spatio-temporal dependencies remains a significant challenge, leading to limited performance gains. Inspired by a recently proposed state space model named Mamba, which has shown remarkable capability of capturing long-range dependency, we propose a new STG forecasting framework named SpoT-Mamba. SpoT-Mamba generates node embeddings by scanning various node-specific walk sequences. Based on the node embeddings, it conducts temporal scans to capture long-range spatio-temporal dependencies. Experimental results on the real-world traffic forecasting dataset demonstrate the effectiveness of SpoT-Mamba.
Paper Structure (20 sections, 2 figures, 3 tables)

This paper contains 20 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. State Space Model (SSM)
  4. Discretized SSM
  5. Mamba
  6. Spatio-Temporal Graph Forecasting
  7. Spatio-Temporal Graph Forecasting with SpoT-Mamba
  8. Multi-way Walk Sequence
  9. Walk Sequence Embedding
  10. Temporal Scan with Mamba Blocks
  11. STG forecasting of SpoT-Mamba
  12. Experiments
  13. Dataset and Experimental Setup
  14. Dataset
  15. Baselines
  16. ...and 5 more sections

Figures (2)

  • Figure 1: Overview of SpoT-Mamba. The first row represents the node-specific walk sequence embedding. The second row represents the overall procedure of STG forecasting. $\mathbf{W} \in \mathbb{R}^{N \times D}$ denotes the node embeddings for all nodes in the graph and $\mathbf{Z}_{t'} \in \mathbb{R}^{N \times 4D}$ denotes one of the outcomes from the temporal scan, corresponding to the time step $t'$.
  • Figure 2: Traffic flow forecasting results for four randomly selected nodes in PEMS04. The blue line represents the ground truth, and the orange line denotes the predictions made by SpoT-Mamba.