Abstract: Existing deep-learning-based methods always divide the predicted region into grids which does not conform to the natural form of accidents,while accidents generally occur on roads. Aiming at the problem of road-level accident risk prediction, an urban traffic accident risk prediction model scale-reduced attention based on graph convolution network (SA-GCN) is proposed. First, the model effectively combines historical long-term and short-term risks, external weather features and a gated graph convolution structure to capture spatial-temporal correlations, and then an attention mechanism is applied to obtain dynamic representations of spatial-temporal features. After that, to solve the problem of the sparseness and spatial heterogeneity of accident data, a scale reduction module which uses the accident risk of the coarse-grained area after clustering, is designed to guide the accident risk prediction at road level. Experimental results on real traffic datasets performance measurement system show that the SA-GCN model performs better than six baseline models, and achieves 11% higher prediction accuracy than the state-of-the-art model.

Key words: graph convolution, attention mechanism, traffic accident risk forecasting