Feasibility Analysis Of Long Short-Term Memory Recurrent Neural Network In Time Series Crime Prediction: A Case Of Bole Sub City Police Department

Abstract

Crime influences people in many ways. Prior studies have shown the relationship between time and crime incidence behavior. This thesis attempts to determine and examine the relationship between time, crime incidences types and locations by using one of the neural network models for time series data that is, long short-term memory network. Six thousand thirty-three (6033) records of five years (2014-2018) data were collected from bole sub city police department in a paper form which has been converted to electronic format. After pre-processing of raw data, it has been analyzed and tested using long short-term memory recurrent neural network model. R-square score is also used to test the accuracy. The result for r-square score on crime type prediction on monthly, daily and hourly basis is 0.879, 0.925 and 0.959 respectively whereas the results on location prediction in monthly, daily and hourly manner are 0.930, 0.989 and 0.968. The study results show that applying long short-term memory recurrent neural network (LSTM RNN) enables to come up with more accurate prediction about crime incidence occurrence with respect to time. Predicting crimes accurately helps to improve crime prevention and decision and advance the justice system.