Researches

A Novel Graph Convolutional Networks to Progressively Adapt to Online Data

​​​​In recent years, graph convolutional networks that possess the ability to adapt to the input data have shown promising results in several studies. In most cases, however, the adaptation has been made during the training phases of the models, which inevitably make the models vulnerable to unexpected traffic conditions such as road closure and traffic accident during the testing phases. In this study, we propose a novel traffic forecasting model, Progressive Graph Convolutional Network (PGCN) to make the model adapt to online traffic data. PGCN constructs a set of graphs by calculating learnable similarity measures among the node signals. The architecture of the model is based on Graph WaveNet. When applied to four real-world datasets, PGCN consistently achieves state-of-the-art performance. This result shows that PGCN has the ability to generalize in different study areas by progressively adapting to online data.

Reference:

• Y. Shin, Y.Yoon*, "PGCN: Progressive Graph Convolutional Networks for Spatial-Temporal Traffic Forecasting", ArXiv preprint, arXiv:2202.08982, 2022 (Under review-IEEE T-ITS). [Link]

Performance Evaluation of Basic Elements of Deep Neural Network Models for Traffic Forecasting

Since 2014, various studies have proposed deep learning-based models to solve traffic forecasting problems. While earlier approaches have shown a more wide range of implementations, the basic elements of the recent models can be categorized into a few groups - RNN, convolution, and self-attention for temporal feature extraction, and Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) for spatial feature extraction. However, there has been little effort to assess the characteristics of each element and make an in-depth evaluation. In this study, we thoroughly evaluated the performance and characteristics of basic elements of traffic forecasting models through extensive and multi-faceted experiments on four real-world datasets. The result reveals that there is no single element that is superior to the others in all aspects. Interesting outcomes include that the convolution-based models show more accurate overall performance than the attention-based models, while the attention-based models show more robustness against abnormal conditions.  

References 

• Y. Shin, Y. Yoon*, "Performance Evaluation of Building Blocks of Spatial-Temporal Deep Learning Models for Traffic Forecasting", (Under review - ESWA). [Link]

• Y. Shin, Y. Yoon*, "A Comparative Study on Basic Elements of Deep Learning Models for Spatial-Temporal Traffic Forecasting", AAAI-22 workshop: AI for Transportation, 2022. [Link]

GNN-based Traffic Forecasting Using Various Structural Characteristics 

Traffic forecasting is a research area in transportation engineering that has garnered broad research interest as a key technical enabler of intelligent transportation systems. The recent surge of Graph Convolutional Networks has enabled traffic forecasting models to incorporate the complex relationship among nodes of transportation networks, and further improved the performance of deep learning-based models. However, many studies overlook the structural characteristics (i.e. speed limit, distance, and flow direction) of transportation networks which can bring important information about the networks. In this research, we propose Multi-Weight Traffic Graph Convolutional (MW-TGC) Networks to incorporate the aforementioned features in a single model and to reflect more spatial correlation in the traffic forecasting problem. Experiments on two real-world datasets show that the proposed MW-TGC Network outperforms the state-of-the-art models. 

References 

• Y. Shin, Y. Yoon*, "Incorporating Dynamicity of Transportation Network With Multi-Weight Traffic Graph Convolutional Network for Traffic Forecasting", IEEE Transactions on Intelligent Transportation Systems, 23(3), 2082-2092, 2022. [Link]