A data mining approach to identify perpetrators: an integration framework and case studies
Data mining and social network analysis have been widely used in law enforcement to solve crimes. Research questions such as strength of ties in social networks, crime pattern discovery and prioritizing offenders have been studied in this area. However, most of those studies failed to consider the noisy nature of the data. The techniques they proposed only have been applied to small scale data sets. Therefore, it is an important task to design a framework that can work on large scale data sets and tolerance noisy data. In this dissertation, we built an integrated crime detection framework that combined two data mining techniques: decision tree and genetic algorithm and graph theories to solve the problems we pointed out. Our crime pattern analysis is based on all offenders of the state of Alabama in the past 50 years. Our constructed social network contains all Alabama residents. It allows us to fully evaluate the proposed models. Two case studies have been conducted to evaluate the framework. One is based on 625 inmates released from Madison county jail in 2004. Our experimental results show that our recommended risk level has strong correlation in predicting future offense. Another case study is based on the 100 real police reports. The experimental results show that the median ranking of arrestees remains at the top 3% of the return list.