Recidivism forecasting: A study on process of feature selection

Abstract

Across the world, there are several factors attributed to high crime rates. The prevention of and the fight against crimes is a major concern of all countries. In the era of globalization and new information and communication technologies, reducing these crimes rate by using conventional methods (law enforcement, social interventions...) are not enough. In fact, they have many limits. Today, by analyzing a large volume of crimes data with machine learning algorithms, researchers can take important advantage of these technologies, especially in the context of the world's famous problem of recidivism. By using these recent innovations, security departments can predict how, when, and where reoffending will happen before it actually happens. However, the efficiency, the quality, and the accuracy of these forcasting models and software depend on several factors. The process of feature selection is one of these key factors. By improving the quality of this process, we can reduce over fitting and eliminating redundant data as well as training time. In this context, this investigation paid particular attention to the process of recidivism features selection (first phase of our future recidivism forcasting framework). Based on detailed study of recidivism theoretical factors, previous and recent methods used in recidivism features selection, we present a comparative study on all key elements used in this phase (features, categories of features and methods of features selection). Our main objective is to prepare an important knowledge database for recidivism features. This database will take into account different sets of recidivism features obtained by all previous and recent projects. It will also be used in our recidivism forcasting framework.