Attentive differential convolutional neural networks for crowd flow prediction

2022 
Traffic crowd flow prediction has drawn more attention in both academic and industry communities due to the explosive growth of traffic data. Generally, existing studies focused on either using as many factors as possible or improving the model structure for describing the spatio-temporal dependence of data. However, due to the complexity and diversity of these influencing factors, few of them can successfully capture the spatial correlation and temporal dependence at the same time. Moreover, none of them have considered the high-order spatio-temporal dependence of traffic data. In fact, mathematically the development of a state can be well described by its current value, together with the series of its high order changing rate. For this insight, in this paper we propose a CNN-based architecture, called Attentive Differential Convolutional Neural Network (ADCNN), to encode the current state and its high order changing rate using the historical traffic data. Inspired by the idea of that a function can be approximated by using a finite number of terms of its Taylor expansion, we design a cascade architecture based on the Hierarchical Differential Unit (HDU) and Cross-Attention mechanism which can learn the high order changing rate of the current state. We conduct extensive experiments to evaluate ADCNN on three real-world traffic datasets. Experimental results show that ADCNN outperforms the other state-of-the-art models.
    • Correction
    • Source
    • Cite
    • Save
    • Machine Reading By IdeaReader
    0
    References
    0
    Citations
    NaN
    KQI
    []