Risk Prediction Model of Recidivism Based on Stacking Algorithm

Abstract

At present, the impact of AI on research in the field of social sciences is gradually attracting attention from the academic community. It has become a trend and made some progress to evaluate the risk of recidivism by quantitative means. However, actuarial assessment, dynamic risk factor assessment and other means take model driven as the basic method, and use human subjective hypothesis to fit the data with unknown distribution pattern, which reduces the objectivity, stability and accuracy of the conclusion to a certain extent, this is the bottleneck which restricting the research in this field. Based on the idea of data-driven, the author uses the dynamic tracking data of 26020 released persons within three years, combined with an ensemble learning method of Stacking, to establish a Intelligent recidivism risk prediction model. This scheme can overcome the limitations of artificial prediction method; mine the potential association relations and mapping rules in the data. At the same time, the author confirms this model is a more reliable and robust method for recidivism prediction than Random Forest model in terms of prediction accuracy, ACU area, elimination of over fitting, small sample learning and other indicators. This model achieves comprehensive performance improvement by adding multiple weak classifiers, so it will be a more efficient prediction method of recidivism based on big data in the future.