Defining Prediction Variables for Theft Crimes by Applying Data Mining Techniques

The prediction and subsequent prevention of theft is an important research topic in criminology, because theft is one of the most common types of crime and it is one that increases fear of crime significantly. There are many possible prediction variables that exist in current socio-economic and demographic datasets; however the form of crime prediction models is unknown, and the relationships between variables are also unclear. For this reason, data mining techniques have been used for variable selection and model building. In this study, we applied decision trees and random forest models to three year-long crime datasets from Taipei City, Taiwan from 2011 to 2014. We found that different theft crimes may have different prediction factors, but several factors exist that are common to many types of theft crimes. For example, areas with a higher density of housing, solitary residents, more CCTV cameras, and more people move out tend to have more burglary theft. We identified several variables that have appeared in prior literature as well as some new variables. As for model performance, data mining techniques did not perform significantly better than traditional regression models. This study is helpful for the police department to define related factors and predict future crime rates, which assists in the design of corresponding policies and enforcement plans to prevent crime.