Graph-ResNets for short-term traffic forecasts in almost unknown cities

Henry Martin, Dominik Bucher, Ye Hong, René Buffat, Christian Rupprecht, Martin RaubalNeurIPS 2019 · OpenReview

TL;DR

This paper introduces Graph-ResNets, a novel graph-based deep learning approach for short-term traffic forecasting. By explicitly incorporating street network information, their models demonstrate superior generalization to previously unseen cities compared to purely image-based methods like CNNs.

What Problem Does This Paper Address?

Urban traffic congestion is a growing problem, leading to pollution, economic costs, and climate change. Accurate short-term traffic flow predictions are crucial for managing traffic efficiently, enabling smart routing, and optimizing transportation systems. However, existing traffic forecasting methods often struggle with standardization, using diverse data resolutions and error metrics, making direct comparisons difficult. Furthermore, many approaches treat traffic data as raw images, neglecting the fundamental constraint that vehicle movement is largely confined to the road network. This paper aims to address this by developing a model that leverages the underlying street network structure for more robust and generalizable traffic predictions, especially in cities where limited training data is available.

Methods

The core of this work is the development of Graph-ResNets, a deep learning architecture that combines graph convolutional networks (GCNs) with residual learning, inspired by ResNet.

  • Graph Extraction: Traffic data, initially in image format, is transformed into a graph. Pixels are treated as nodes, connected to adjacent pixels via undirected edges. To reduce computational complexity and focus on active traffic areas, nodes corresponding to low-activity pixels are pruned. Node features are derived from the traffic data channels (volume, speed, direction).
  • Graph Convolutional Networks (GCNs): The paper utilizes GCNs, which are designed to operate on graph-structured data. They employ a standard GCN block consisting of graph convolution and ReLU activation.
  • Skip Connections (SkipfNet): To improve the ability of GCNs to learn identity-like functions and prevent over-smoothing in deeper networks, skip connections are introduced. This allows the input of a layer to be concatenated with its output.
  • Graph Residual Networks (Graph-ResNet): Building upon the SkipfNet, the Graph-ResNet incorporates residual blocks. This means the output of a graph residual block is the sum of the input and the processed output, further enhancing the network's ability to learn complex functions and enabling deeper architectures without performance degradation.

The authors compare their Graph-ResNet and SkipfNet models against U-Net (a CNN-based architecture) and the original KipfNet GCN.

Datasets

The primary dataset used is the Traffic4cast dataset, provided by the Institute of Advanced Research in Artificial Intelligence (IARAI). This dataset comprises traffic data from three cities: Berlin, Istanbul, and Moscow. The data is provided as three-channel images representing traffic volume, average speed, and average direction, with 5-minute intervals over one year. The authors train their models on Moscow data and then evaluate their generalization capabilities on Berlin and Istanbul.

Key Insights and Ideas

  • Street Network as Explicit Prior: The most significant insight is that explicitly providing the street network structure to a deep learning model (via graph representation) leads to better generalization to unseen cities. This contrasts with CNNs, which must implicitly learn this structure from image data.
  • GCNs Generalize Better to Unknown Cities: While CNNs (specifically U-Nets) perform well on the city they were trained on (Moscow), GCN-based models, particularly Graph-ResNets, show superior performance when predicting traffic in different, unseen cities (Berlin and Istanbul).
  • Residual Connections are Key for Deeper GCNs: Similar to image-based ResNets, residual connections are crucial for enabling deeper GCN architectures without suffering from over-smoothing or performance degradation.
  • Trade-off in Competition Design: The image-based nature of the Traffic4cast competition and its evaluation metric (pixel-wise MSE) inherently favors CNNs. GCNs, by design, can only predict values on the graph (i.e., on roads), which can lead to lower scores under these specific competition rules, even if the underlying prediction quality on the road network is superior.

Key Results

  • Generalization: Graph-based models (Graph-ResNet, SkipfNet) consistently outperformed CNN-based models (U-Nets) in predicting traffic in unseen cities (Berlin and Istanbul) when trained on Moscow data.
  • Moscow Performance: For predicting traffic in Moscow (the training city), deeper U-Net models achieved slightly better results than the graph-based models, suggesting CNNs are highly effective when the target domain is known and well-represented in training.
  • Graph-ResNet Superiority: Among the graph-based models, the Graph-ResNet emerged as the best performer, indicating the benefit of residual connections for deeper GCN architectures.
  • Parameter vs. Performance: For GCNs, a correlation between the number of parameters and performance was observed, with more complex models generally performing better. This relationship was less clear for deeper U-Nets, potentially due to training issues.

Limitations and Future Work

  • Competition Rules Disadvantage: The authors acknowledge that the competition's evaluation metric (pixel-wise MSE) and the inclusion of non-road pixels disadvantage graph-based approaches. Future work could involve evaluating models using metrics that specifically focus on road pixels.
  • Low Threshold for Graph Extraction: To perform well under competition rules, a low threshold was used for extracting the street network, which almost reproduces a grid graph. This can diminish the advantage of GCNs over CNNs.
  • Graph Pooling: The integration of graph pooling methods is identified as a key area for future research. Graph pooling is complex but could allow for more efficient and scalable graph network architectures, potentially leading to further performance improvements, especially for larger graphs.
  • Exploring More Complex Graph Models: Further investigation into more sophisticated graph model architectures and their relationship with performance on unknown cities is suggested.

Summary updated

Abstract

The 2019 IARAI traffic4cast competition is a traffic forecasting problem based on traffic data from three cities that are encoded as images. We developed a ResNet-inspired graph convolutional neura...
BibTeX
@inproceedings{martin2019graph,
  author = {Henry Martin and Dominik Bucher and Ye Hong and René Buffat and Christian Rupprecht and Martin Raubal},
  title = {{Graph-ResNets for short-term traffic forecasts in almost unknown cities}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=9Kc1cxkjUIi},
}
Figure