NCJ Number
177224
Date Published
1998
Length
98 pages
Annotation
Automated methods of proactive pattern recognition are not currently available to police crime analysts; reactive methods exist in the form of database queries and geographic information systems, but these methods are not adequate for proactive pattern recognition.
Abstract
Police detectives in large urban police departments are not likely to proactively discriminate crime patterns due to human information processing constraints and large crime data sets. The volume of criminal cases reported and the number of characteristics of each criminal incident make it virtually impossible to match and compare the extensive data involved over time. In order to overcome information processing deficiencies and identify at least some portion of pattern criminal incidents, police detectives construct various heuristics or decision shortcuts. For example, since vehicle descriptions are rarely reported in robbery cases, some police detectives will look for cases with a vehicle described and will then search only this variable in other cases. Given there will be very few cases with this variable described, the search for patterned criminal incidents where the offender used a particular vehicle will be relatively manageable. Heuristic construction is related to various organizational structures, and institutional arrangements of an organization will affect incentives and disincentives that encourage or discourage police detectives to construct heuristics. In particular, specialization of the police detective function will likely serve to facilitate pattern recognition heuristics. A project involving police robbery detectives in the Chicago Police Department attempted to create an automated crime pattern recognition tool to classify patterns of similar criminal cases using a neural network approach. Employing supervised learning with undifferentiated crime patterns resulted in suboptimum performance of the neural net. There were no visible patterns to be used as training patterns, and the back propagation network architecture was not able to differentiate patterns of criminal activity given the data available. The nearest neighbors approach exhibited greater promise than the neural network approach in clustering like criminal cases. A 19- dimension Euclidean space was constructed, values were assigned to each variable, and the nearest neighbor clustering method was used to determine the relative spatial positioning of criminal cases. Initial cluster validation indicated the output was tactically useful to police detectives and led to crime pattern classification. An appendix contains information on files, codes, and categories used in crime pattern analysis. 63 references and 2 figures