Correlated Channeled Spatio Temporal Graph Attention Network Model for Traffic Prediction

Accurate real-time traffic prediction in intelligent transportation systems, remains a challenging task because of the frequently changing spatio-temporal dependencies, complex dynamics of the network and varying traffic density. While, existing methods have widely explored the hidden linkages for real-time traffic prediction, multi-faceted interactions between various traffic signals can play an important role in improving the accuracy rate of such systems. This work presents a novel method for analyzing the interrelated interactions between different features of traffic signals through a co-related channeled-spatio-temporal graph attention network. The proposed model is aimed to reflect the intricate channel, temporal and spatial connections between various traffic signal components. The model learns three different embeddings to describe the interactions between traffic signals using graph attention networks. These embeddings are in-line with the signals temporal dynamics, spatial interactions, and channel-based features. Correlation score is used to evaluate the degree of similarity between nodes in various time windows, which ultimately aids in the selection of the most important data. The proposed model is able to outperform various state-of-the-art methods in a series of trials on five real-world datasets. A comprehensive analysis of the results prove that the proposed model not only captures the spatio-temporal correlations in traffic patterns, but is also able to take into account the interactions between traffic signals and their impact on traffic flow changes over time, leading to a 9.7% improved error rate.