An Associative Memory for Predicting Crime
|has title::Geopredict:Geographical Crime Forecasting for Varying Situations|
|Master:||project within::Computational Intelligence and Selforganisation|
|Student name:||student name::Bas Weitjens|
|Second reader:||has second reader::Selmar Smit|
|Company:||has company::Sentient Information Systems|
The efficiency of the police would greatly improve if they would be able to predict crime. To help the police get a better insight many different crime forecasting techniques have been developed. Among these techniques are geographical crime forecasting techniques, to create forecasts at what location crime is likely to take place. In this paper different geographical crime forecasting models are compared. The models selected for the comparison are: two random walk models, an ARIMA model with a spatial extension and two DataDetective models. The two random walk models differ in what data they use for the forecasts, and the two DataDetective models differ in the way their settings are optimized. Different measures are taken to make the comparison between the techniques fair. To calculate the performance as accurate as possible two different error measurements are used: the NRMSE after smoothing, and the search efficiency. Also the parameters of each of the techniques are optimized to make sure possible differences in performance are not caused by parameter settings. All of the techniques are compared on varying situations, to see the effect of multiple aspects of a situation. For the optimization of the DataDetective technique an evolution strategy is used. For one of the two DataDetective models the optimization process is simplified, among others by using trend functions, while for the other model extensive optimization is used. The two different types of optimization are used to see the difference in performance compared to the calculation time needed. The results of the optimization show that the parameters of the DataDetective technique do not have a large impact on the performance, and therefore the gains from optimization are small. The results of the comparisons between the different models show that ARIMA performs worst. The two DataDetective models perform slightly better than the random walk models. Between the two DataDetective models the one with extensive optimization performs somewhat better, but the differences in performance are often very small. From these results it is concluded that it is good practice for the police to use the DataDetective technique with simplified optimization for situations that are used relatively seldom, while using the DataDetective technique with extensive optimization for situations that are used often, since it is worth the extra calculation time in those cases.