Traffic Graph Convolutional Recurrent Neural Network: A Deep Learning Framework for Network-Scale Traffic Learning and Forecasting

Cui, Zhiyong; Henrickson, Kristian; Ke, Ruimin; Wang, Yinhai · 2019 · OpenAlex-citations

DOI: 10.1109/tits.2019.2950416

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of network-scale traffic forecasting, which requires modeling complex spatiotemporal dependencies and time-varying traffic patterns. Existing methods, including statistical models and standard deep learning approaches like Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), often fail to adequately capture the physical topology of road networks or interpret spatial relationships. Specifically, conventional CNNs are ill-suited for non-Euclidean graph structures, while spectral graph convolution methods often lack interpretability and neglect physical roadway characteristics. To resolve these limitations, the authors propose a novel deep learning framework called Traffic Graph Convolutional Long Short-Term Memory (TGC-LSTM). The methodology centers on modeling the traffic network as a graph where nodes represent sensing locations and edges represent connections. The core innovation is a Traffic Graph Convolution (TGC) operator that incorporates physical network properties, specifically using a "free-flow reachable matrix" to define which nodes can influence each other within a given time interval based on free-flow speeds and distances. This approach confines the receptive field to physically relevant neighborhoods, unlike standard graph convolutions that may capture spurious long-range dependencies. The TGC extracts localized spatial features from multiple hop orders, which are then fed into an LSTM layer to model temporal dynamics. The LSTM architecture is modified to include a cell state gate that accounts for the influence of neighboring cell states, constrained by the same physical reachability matrix. Additionally, L1 and L2 regularization terms are added to the loss function to enhance the interpretability and stability of the learned graph convolution weights. Experimental results on two real-world traffic state datasets demonstrate that the TGC-LSTM model outperforms multiple state-of-the-art baseline methods in forecasting accuracy. The study highlights that the model’s ability to incorporate physical specialties of traffic networks, such as road length and free-flow speed, allows it to more accurately capture the true causal structure of traffic impact transmission. Furthermore, the visualization of the learned graph convolution weights reveals that the model can successfully identify the most influential roadway segments in real-world networks. This interpretability distinguishes the proposed framework from black-box alternatives, providing insights into spatial dependencies that align with traffic flow theory. The significance of this work lies in its integration of physical traffic constraints into deep learning architectures, bridging the gap between data-driven models and domain-specific knowledge. By defining convolution operations based on free-flow reachability rather than arbitrary graph topology, the TGC-LSTM offers a more robust and interpretable solution for network-wide traffic prediction. The authors also contribute to the field by publishing the real-world traffic speed data and graph structures used in the study, facilitating further research. This framework represents a step forward in developing intelligent transportation systems that can effectively leverage large-scale traffic data while respecting the underlying physics of road networks.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• mental model of traffic