Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GraphSparseNet: a Novel Method for Large Scale Traffic Flow Prediction

Weiyang Kong, Kaiqi Wu, Sen Zhang, Yubao Liu

TL;DR

This work tackles the scalability of GNN-based traffic forecasting by introducing GraphSparseNet (GSNet), a two-module framework consisting of a Feature Extractor and a Relational Compressor that operate in low-dimensional spaces to achieve linear time and space complexity. The authors provide a theoretical basis for replacing a full adaptive adjacency with a low-rank, coefficient-guided representation (via matrices K and U) and demonstrate how both modules collaborate to fuse node features and model sparse relationships efficiently. Empirically, GSNet delivers competitive or superior predictive accuracy across four real-world datasets while significantly reducing training time (up to 3.51× faster than state-of-the-art linear models) and maintaining low memory usage, particularly on large-scale networks. These results suggest GSNet offers a practical, scalable solution for large-scale traffic flow forecasting and related spatio-temporal prediction tasks.

Abstract

Traffic flow forecasting is a critical spatio-temporal data mining task with wide-ranging applications in intelligent route planning and dynamic traffic management. Recent advancements in deep learning, particularly through Graph Neural Networks (GNNs), have significantly enhanced the accuracy of these forecasts by capturing complex spatio-temporal dynamics. However, the scalability of GNNs remains a challenge due to their exponential growth in model complexity with increasing nodes in the graph. Existing methods to address this issue, including sparsification, decomposition, and kernel-based approaches, either do not fully resolve the complexity issue or risk compromising predictive accuracy. This paper introduces GraphSparseNet (GSNet), a novel framework designed to improve both the scalability and accuracy of GNN-based traffic forecasting models. GraphSparseNet is comprised of two core modules: the Feature Extractor and the Relational Compressor. These modules operate with linear time and space complexity, thereby reducing the overall computational complexity of the model to a linear scale. Our extensive experiments on multiple real-world datasets demonstrate that GraphSparseNet not only significantly reduces training time by 3.51x compared to state-of-the-art linear models but also maintains high predictive performance.

GraphSparseNet: a Novel Method for Large Scale Traffic Flow Prediction

TL;DR

This work tackles the scalability of GNN-based traffic forecasting by introducing GraphSparseNet (GSNet), a two-module framework consisting of a Feature Extractor and a Relational Compressor that operate in low-dimensional spaces to achieve linear time and space complexity. The authors provide a theoretical basis for replacing a full adaptive adjacency with a low-rank, coefficient-guided representation (via matrices K and U) and demonstrate how both modules collaborate to fuse node features and model sparse relationships efficiently. Empirically, GSNet delivers competitive or superior predictive accuracy across four real-world datasets while significantly reducing training time (up to 3.51× faster than state-of-the-art linear models) and maintaining low memory usage, particularly on large-scale networks. These results suggest GSNet offers a practical, scalable solution for large-scale traffic flow forecasting and related spatio-temporal prediction tasks.

Abstract

Traffic flow forecasting is a critical spatio-temporal data mining task with wide-ranging applications in intelligent route planning and dynamic traffic management. Recent advancements in deep learning, particularly through Graph Neural Networks (GNNs), have significantly enhanced the accuracy of these forecasts by capturing complex spatio-temporal dynamics. However, the scalability of GNNs remains a challenge due to their exponential growth in model complexity with increasing nodes in the graph. Existing methods to address this issue, including sparsification, decomposition, and kernel-based approaches, either do not fully resolve the complexity issue or risk compromising predictive accuracy. This paper introduces GraphSparseNet (GSNet), a novel framework designed to improve both the scalability and accuracy of GNN-based traffic forecasting models. GraphSparseNet is comprised of two core modules: the Feature Extractor and the Relational Compressor. These modules operate with linear time and space complexity, thereby reducing the overall computational complexity of the model to a linear scale. Our extensive experiments on multiple real-world datasets demonstrate that GraphSparseNet not only significantly reduces training time by 3.51x compared to state-of-the-art linear models but also maintains high predictive performance.

Paper Structure

This paper contains 22 sections, 1 theorem, 35 equations, 9 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Traffic Network
  4. Problem Definition
  5. Analysis
  6. Proposed Model
  7. Framework of the model
  8. Feature Extractor
  9. Relational Compressor
  10. Complexity Analysis
  11. Experments
  12. Datasets and Pre-processing
  13. Baselines
  14. Experimental Setup
  15. Comparison with Existing Models
  16. ...and 7 more sections

Key Result

theorem 1

Let $M \in \mathbb{R}^{N \times N}$ be a matrix with rank $C$. There always exists a non-unique matrix $K \in \mathbb{R}^{C \times C}$ such that matrix $M$ can be constructed via some matrix multiplication transformations involving $K$.

Figures (9)

  • Figure 1: Visualizations correspond to the well-trained adjacency matrix obtained from existing graph neural network method. It can be observed that matrix is highly sparse. Statistical data on the weighted degree also demonstrates the sparsity of the adjacency relationships.
  • Figure 2: An illustration of the framework.
  • Figure 3: An illustration in Relational Compressor.
  • Figure 4: Memory usage on all the datasets.
  • Figure 5: Prediction performance of different variants on PEMS07 and CA datasets.
  • ...and 4 more figures

Theorems & Definitions (1)

  • theorem 1