Hotspot Prediction Based on Temporal Characteristics

Nowadays, with the expanding population and city size, urban crime rate control will be a very important direction for the integration of artificial intelligence and urban police governance, and the prediction of the number of regional hotspots is an effective crime prevention method. Based on the real dataset of a city, we hope to improve the prediction effect of crime hotspots and make analysis and prediction feasibility judgments from the perspective of data analysis. First, the dataset was pre-processed and filtered, and then analyzed from the spatial and temporal perspectives to further judge the feasibility of prediction. From a non-spatial perspective, the effect of adding covariates on the prediction of urban crime hotspots was explored. First, a map was drawn based on the distribution of hotspots, the map was divided into a grid, and the grid information was classified into four categories by clustering, and then selected covariates were added to the model for experiments. This study extends the prediction range in terms of temporal characteristics. The differential integrated moving average autoregressive model (ARIMA) is commonly used for time series forecasting, but it is more suitable for dealing with linear data. The long short-term memory neural network (LSTM) has a strong advantage in dealing with nonlinear data. We construct a combined ARIMA-LSTM model. It can fully exploit the data information and improve prediction accuracy. The results show that the combined ARIMA-LSTM model can predict the property crime in a district of the city better than the single ARIMA model and LSTM model, and the combined model can better fit the actual trend of the cases.