Urban Traffic Accident Risk Prediction Based On Graph Neural Network

As people’s travel needs and the number of urban vehicles grow,the number of traffic accidents is also rising year by year.Vicious traffic accidents can seriously endanger the lives and property safety of urban residents.Therefore,if we can establish an appropriate model,make a reasonable prediction of the risk of traffic accidents in the city,and inform the city residents and relevant management departments of the results of the prediction in advance,it will effectively improve the traffic situation in the city and reduce people’s travel risks.Urban traffic risk prediction is a typical spatio-temporal sequence prediction problem,and in the past,due to the lack of data and methods,people have been slow to study on this problem.In recent years,with the development of big data and deep learning networks,some work has begun to use deep learning to solve the problem of urban traffic risk prediction,and the prediction results have been improved to a certain extent.However,traditional convolutional neural networks and recurrent neural networks do not deal well with spatiotemporal problems,and there is little work that can deeply consider the intrinsic relationship between traffic flow changes and traffic accidents.To this end,this paper proposes a new urban traffic accident risk prediction model based on graph neural network.Compared with previous models of accident risk prediction,this model takes into account the intrinsic relationship between regional traffic changes and accidents at a deeper level.At the same time,this paper adopts the construction strategy of dynamic graph adjacency matrix,which omits the cumbersome process of manually constructing static graph adjacency matrix of traditional graph neural network,improves universality,and dynamically reflects the traffic flow changes in different sub-regions of the city,so that the model’s perception of traffic changes is improved.At the same time,by analyzing the historical data of traffic risk,this paper finds a relatively serious double data imbalance in the dataset.The first data imbalance due to the small number of sub-regions where accidents occurred in the same time period and the second data imbalance due to the distribution of most accidents in a small number of accident-at-risk areas in the global time period.In this paper,the label replacement strategy based on risk a priori and the loss function weight adjustment strategy based on high-risk area optimization are proposed,which improves and alleviates the imbalance between the two data to a certain extent,so that the final prediction effect is improved.Finally,the model comparison experiment and ablation experiment are carried out on the real data set,which is improved compared with the traditional machine learning and deep learning models,which proves the superiority of the proposed model and the effectiveness of the two strategies to solve the double imbalance.