The purpose of police crime prediction is to directly support crime prevention and law enforcement. It is an exciting new area, which brings together the disciplines of statistics, machine learning, artificial intelligence, criminology, and psychology and database technology.
The prediction of crime has seen an increase of research attention over the last decades and is more widely practiced by police. This is due to fact that Geographical Information systems (GIS) had become an important tool for police agencies. The mapping of crimes and the identification of hot spots has become regular practice. The ‘criminality profile’ of places was established based on theories like routine activities, the ecology of crime and hot spots.
There have been numerous studies using varity of related methodology such as statistic, neural networks for the spatial and temporal analysis of crime; the following outlines some of the recent ones. Law enforcement agencies have a continuing need to predict time and locations of crimes. Liu and Brown (2003) showed that predicted crime locations for next week based on data from the previous week. They suggest that the likelihood of a criminal incident at a specified location is based on past incidents of the same type and independent spatial attributes or features. They compare two prediction models for hot spots that relate the features in an area to the predicted occurrence of crime through the preference structure of criminals, and conclude that the model performs significantly better with the extra features.
Gorr and Olligschlaeger (2003) study monthly crime data over the period 1991- 1998. Using holdout samples the subset of actual time and rolling horizon experimental design, to compare the forecasting accuracy of model fit to past data. They contrast the forecast accuracy of univariate time series models with naïve methods. This method used time series data points themselves as forecasts. They find the classical seasonal patterns of increased property crime levels late in the year and increased aggression crimes in summer due to increased social interactions, cold weather might increase burglary, and robbery crimes due to seasonal economic pressures or unemployment. Felson and Poulsen (2003) studies show that crime varies greatly by hour of day more than by any other variable and criminal activity
between 5: AM and 4:49 AM the next morning. This suggests that a prediction for specific time periods might be valuable for police planning.
Several studies show risk factors of crime. Roncek and Maier (1991) found relationship between levels of crime and the number of taverns and lounges located. Drug hotspots tended to be in areas with poverty and low family cohesion Gorr and Olligschlaeger (1993).
Bowers et al. (2004) investigated the relationship between area type, housing type, level of victimization and repeat victimization. The results have demonstrated that the influences of area and housing type being burgled interact. For instance a detached house located in deprived area is at over seven times the risk of a detached house in affluent areas. Analysis of the relationships between the spatial patterns of residential burglaries and the socio economic characteristics of neighborhoods in London has been examined by Malczewski, et al. (2005) using geographically weighted regression. The result shows that there were significant spatial variations in the relationships between the relative risks of residential burglaries and the average value of dwelling and the percentage of multifamily housing. Edmark (2005) studied the effects of unemployment on property crime rates. According to the theory of economics of crime, increased unemployment rates lead to higher property crime rates. A high crime rate leads to unemployment because new firms do not want to settle in a criminal area and existing businesses leave. It might also be the case that people who have once been in prison have difficulties finding a job and for this reason contribute to a high unemployment rate.
ANN’s are presented as one technique that offers minimal user interaction in addition to dynamic adaptability, and thus a potential operational forecasting solution. One of the earliest was that of Olligschlaeger (1997), who employed back- propagation to predict areas where future drug markets will emerge. In more recent work, Olligschlaege and Gorr (2001) found that ANNs outperform multiple
regression leading indicator models when the set of leading indicators is rich and numerous.
Corcoran, et al. (2003) used hotspots (spatial clusters of crime) for forecasting. The Gamma Test (GT) was applied to each cluster to assess suitability for predictive modelling. ANN and comparative linear regression forecasting models were constructed using the GT, and compared to a “random walk” model. For crime analysis software, Oatley and Ewart (2003) used a Kohonen neural network for matching crimes against the offender list and the Bayesian belief network for prediction of re-victimisation.
Craglia, et al. (2001) reported the strengths of GIS based spatial analysis with census (socioeconomic) data for modelling high-intensity urban crime area. Three police force areas in England and Wales were used to develop the model. These areas that raise special policing problems, such as sometimes found violent forms of crime within them, resident population defect to co-operate with the police. The model suggests that high-intensity crime areas are characterized by populations that are deprived and live at high density and have higher levels of population turnover. This is done through a statistical analysis (regression) which uses data from the census. The spatial datasets within the GIS was used to integrate data on the boundaries of the high-intensity area with socio-economic data. Ratcliffe (2001) derived his study over 14,000 burglaries over two years for separately examining the spatial and temporal patterns of residential and non residential burglary. The study showed that the highest probability for residential burglaries was between 8am and about 6pm, the period that most people were at work. The residential burglary levels were lower over the weekend and overnight. For the spatial analysis of residential burglary the researchers examined the Canberra region. ‘Hotspots’ include the more established suburbs of the inner-north of the city and the inner south-east. The housing characteristics of the residential burglary hotspots vary considerably across the city.
This thesis presented the developing a hybrid predictive models for crime based on real data (burglary incidence). Both regression methodology and neural networks have been used for predictive crime modelling. Historical data with background population and census datasets are used for predictive crime modelling. The obtained models based on the observed data in the study region and which presented in chapter six are reasonable.