With the continuous increase in the number of road vehicles, traffic accidents have become one of the main factors that endanger social public safety, and the prediction of road traffic accidents has also become a research hotspot.Taking into account the intricacies of accident influencing factors and the dynamic spatio-temporal variability and data sparseness of accidents, the fusion of multi-source data and the feature extraction according to time-varying and time-invariant data, especially the text description of the accident The feature extracts context information, and at the same time, the negative sampling method is used to balance the ratio of positive and negative samples.Finally, a regional traffic accident prediction network model trained by multi-feature component combination training is proposed. The model was validated on the data sets of three cities with different accident sparsity in the United States. The experimental results show that the prediction model is better than the basic model of comparison in various evaluation indicators, and each indicator is increased by about 2%3%. It can be seen that the model has improved the prediction performance to a certain extent. At the same time, the experimental results of different combinations of multi-feature components show that various factors have an impact on the occurrence of accidents.