Using data-mining and multi-agent simulation to predict criminal behavior

In this work, the events caused by criminals with similar preferences are modeled as the events caused by agents. The behavior of these agents is used to create the predictive model. However, some method must be used to partition the events caused by the agents. In order to create this partitioning method, an estimation of the types of distributions of the events caused by criminals must be made. Judgment analysis, a descriptive decision theory, is used to make this estimation. If all of the agents are using the same set of features to select specific targets, then a density based clustering method will partition the events caused by the agents. A set of independent agents was discovered from the data on the objects stolen in Richmond. Association rules and linear regression were used to suggest that thieves in Richmond are not using the same sets of features to select specific targets. This means that some other method to partition them must be devised.; The cluster specific salience discovery (CSSD) methodology was created to partition the agents, where the agents are using different sets of features to select targets. This methodology is used with existing clustering methods. In this methodology: (1) A variance threshold is discovered for each feature; (2) A minimum number of points for a cluster is selected; (3) The events are clustered in the space formed by all the feature; (4) Clusters that have a variance of less than the variance threshold are removed; (5) The remaining events are clustered in the space formed by a subset of the features; (6) This continues until either the number of points falls below the minimum number of points or all of the feature subspaces have been explored.; This methodology was tested on synthetic data with two clustering methods. The use of the CSSD methodology with two clustering methods provided a statistically significant improvement over the same clustering methods used without the methodology.; In the Richmond data set, only one variable, the average temperature of the previous day (Lagged Temperature), changes over the period of study. In order to create a predictive model, a generalized linear model with a Poisson distribution, was created for each of the discovered agents. For some of the agents, a predictive model could be created using the Lagged Temperature variable, and the number of crimes committed on the two previous days. The predictive model for those agents provides a statistically significant improvement over other methods.; The following contributions are made by this work: (1) A model of criminal behavior based on judgment analysis. This model allowed for the construction of a model of criminal behavior that includes the possibility of agents using different sets of features. (2) A demonstration of the use of different features sets by criminals in the target selection process. (Abstract shortened by UMI.)...