A Hybrid GCN and LSTM Structure Based on Attention Mechanism for Crime Prediction

Abstract

Globalization has been the major contributor to economy boom. While at the same time, it has stimulated the development of crime method as frequent cross-border communication allowed. With the improvement in big data and prediction system of policing work, it has become a new research field to establish an efficient crime prediction model, by which police departments could clamp down on criminal activities more accurately. Besides, this model will be quite beneficial for commanding and dispatching police force thus to improve work efficiency. This paper proposes a combination model, which uses Long Short-Term Memory Network (LSTM) and Graph Convolutional Network (GCN) to predict crime rate and takes advantage of the Attention mechanism to improve the experimental result. By extracting the spatio-temporal characteristics of crimes and increasing the proportion of typical feature, it can not only predict crime quantity, but also detect the degree of crime risk in each region. A rolling forecast of crime data for about three years in the Boston of the United States shows that our model has good prediction performance.