Revisitng graph neural networks for traffic forecasting
Keywords: Graph-structured dynamics, Linear model, Large network, Traffic forecasting
TL;DR: A simple and effective benchmark method for traffic forecasting
Abstract: Accurate traffic forecasting is crucial for a wide range of traffic management applications. In recent years, Graph Neural Networks (GNNs) have emerged as one of the most promising methods to predict traffic. However, their complex architectures prevent them from being used in large networks and long-term forecasting. Although there are complex spatial-temporal dependencies in traffic data, the evolution of the traffic flow, in particular, is governed by linear dynamics, based on the law of flow conservation. Hence, we conjecture that linear models are sufficient for accurate traffic flow predictions. In this study, we investigate linear regression models to predict traffic flow. Models are created for different periods in the day, and exploit historical traffic data from the neighboring region as input. Using multiple real-world traffic data sets collected from the entire California highway systems, we demonstrate that our simple linear models outperform state-of-the-art GNNs by achieving both higher accuracy and significantly better efficiency. Moreover, we conduct comprehensive studies to analyze the impacts of various design elements of GNNs on the improvement of prediction accuracy. Based on our findings, we advocate re-considering the design of model architectures for traffic forecasting.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8523
Loading