Spatio-temporal crash severity analysis with cost-sensitive multi-graphs attention network

Most conventional crash severity models attempt to achieve a low classification error rate, implicitly assuming the same losses for all classification errors. In this paper, we suggest that this setting has limitations in terms of reasonableness, as accurately identifying the significant contributing factors to more severe crashes is more important than identifying those for less severe crashes in crash severity analysis. In addition, the spatio-temporal heterogeneity in existing crash severity models is usually explored utilising statistical models or pre-learned by shallow machine learning models. To the authors' best knowledge, the advanced deep neural networks have received less attention in the context of learning spatio-temporal crash severity structure. To tackle these two issues, this paper firstly reformulates the crash severity analysis as a cost-sensitive learning problem, where a cost matrix is defined to measure the unequal misclassification losses in crash severity analysis. Furthermore, to obtain a more accurate representation of spatio-temporal heterogeneity for crash severity analysis, we develop the multi-graphs attention mechanism, which takes the advanced graph convolutional network as the base learner. As a result, a cost-sensitive multi-graphs attention network (CSmGAT) model is finally proposed, which, on the one hand, learns the optimal spatio-temporal affiliations from crash severity data by performing multi-graphs attention convolutions to filter out the error affiliations in pre-defined spatio-temporal graphs and, on the other hand, embeds the cost matrix and aims to perform cost-sensitive crash severity analysis by minimising the overall misclassification losses of training data. In the experimental part, a five-year vehicle-vehicle crash dataset from Victoria State, Australia is utilised for crash severity analysis. In comparison with 23 state-of-the-art crash severity models, our proposed CSmGAT model can reduce the overall misclassification losses by at least 11.31%. The most significant crash contributing factors have also been identified and interpreted based on the pseudo-elasticity values.