# 3D Graph Convolutional Networks with Temporal Graphs: A Spatial Information Free Framework For Traffic Forecasting

**Authors:** Bing Yu, Mengzhang Li, Jiyong Zhang, Zhanxing Zhu  
**arXiv:** [1903.00919](https://arxiv.org/abs/1903.00919)  
**Date:** 2019-03-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1903.00919v1/

## Key points

- Spatio-temporal prediction is crucial in traffic forecasting for various applications.
- Existing methods struggle with heavy training costs and accurately capturing spatio-temporal patterns and correlations between distant roads.
- The 3D Temporal Graph Convolutional Networks (3D-TGCN) model, proposed by Bing Yu et al., addresses these issues.
- The model learns road graph based on comparing the similarity between time series data for each road, rather than spatial information.
- An original 3D graph convolution model is introduced to better capture spatio-temporal data dynamics.
- Empirical results show that the 3D-TGCN model outperforms state-of-the-art baselines in traffic forecasting tasks.

## Summary

In the field of traffic forecasting, spatio-temporal prediction is crucial for various applications. However, it remains challenging due to complex dependencies and nonlinear dynamics in road networks. Existing methods often struggle with heavy training costs or fail to accurately capture spatio-temporal patterns and correlations between distant roads with similar traffic patterns. To address these issues, a novel deep learning framework called 3D Temporal Graph Convolutional Networks (3D-TGCN) is proposed in this paper by Bing Yu, Mengzhang Li, Jiyong Zhang, and Zhanxing Zhu. The key innovation of the 3D-TGCN model lies in two main components. Firstly, instead of constructing the road graph based on spatial information, the model learns it by comparing the similarity between time series data for each road. This approach provides a spatial information-free framework that enhances the accuracy of traffic predictions. Secondly, an original 3D graph convolution model is introduced to better capture the spatio-temporal data dynamics. Empirical results demonstrate that the 3D-TGCN model outperforms state-of-the-art baselines in traffic forecasting tasks. By effectively modeling spatio-temporal relationships and leveraging temporal graphs without relying on spatial information, this innovative framework shows promise for improving prediction accuracy and efficiency in traffic domain applications. The research conducted by Yu et al. opens up new possibilities for enhancing traffic forecasting models through advanced deep learning techniques tailored to address the complexities of road network dynamics.