A dynamic graph convolutional network with multiscaled attention for traffic prediction
TL;DR
This paper aims to propose attention-based dynamic graph convolutional recurrent neural network (ADGCRNN) for highway traffic flow prediction, which outperforms state-of-the-art baseline models and realizes multiresolution temporal fusion via self-attention.
Abstract
Traffic flow prediction is vital for highway road planning and congestion alleviation. Nevertheless, highway traffic forecasting faces difficulties caused by complex spatio-temporal properties. Existing graph-convolution-based prediction approaches cannot sustain stable spatio-temporal consistency for long horizons. They neglect dynamic spatio-temporal correlations, convolution locality and multiresolution long-term dependencies, limiting prediction accuracy. This paper aims to propose attention-based dynamic graph convolutional recurrent neural network (ADGCRNN) for highway traffic flow prediction. This work presents the ADGCRNN. Self-attention integrates three-resolution temporal sequences for feature extraction. Dynamically constructed multidynamic graphs and adaptive weights capture variant traffic properties. A gated kernel focusing on highly correlated nodes is adopted on full graphs to mitigate graph-convolution overfitting. Evaluated on two public data sets, the proposed ADGCRNN outperforms state-of-the-art baseline models. A practical case study based on a real-world web system further validates the practical benefits of this approach for highway-transportation scenarios. This model realizes multiresolution temporal fusion via self-attention. It leverages adaptive multidynamic graphs to model time-varying spatial patterns. A gated kernel is introduced to alleviate overfitting for full-graph convolution in traffic forecasting.
