Data-driven models in machine learning for crime prediction

Prediction of future events over time is associated with a sequence of time series observational samples and other exogenous data. Different approaches connected with statistical learning techniques result in predictive databased models. The paper presents an attempt to develop techniques for predictive data-based modeling based on machine learning data-driven approaches. To reach a good level of prediction we use a deep learning architecture based on artificial neural network (ANN). The neural network (NN) structure for crime prediction and the appropriate inputs for crime prediction is performed through: Gram-Schmidt orthogonalization (GS) for the selection of network inputs and virtual leave-one-out test (VLOO) for the selection of the optimal number of hidden neurons. Spatiotemporal distribution of the hot-spots is conducted and a methodology is developed for short-term crime forecasting using the long short-term memory (LSTM) recurrent neural networks (RNN) and convolutional neural networks (CNN).