A Predictive Analytics-Based Decision Support System for Drug Courts

摘要

This study employs predictive analytics to develop a decision support system for the prediction of recidivism in drug courts. Based on the input from subject matter experts, recidivism is defined as the violation of the treatment program requirements within three years after admission. We use two data processing methods to improve the accuracy of predictions: synthetic minority oversampling and survival data mining. The former creates a balanced data set and the latter boosts the model's performance by adding several new, informative variables to the data set. After running several tree-based machine learning algorithms on the input data, random forest achieved the best performance (AUROC = 0.884, accuracy = 80.76%). Compared with the original data, oversampling and survival data mining increased AUROC by 0.068 and 0.018, respectively. Their combined contribution to AUROC was 0.088. We present a simplified version of decision rules and explain how the decision support system can be deployed. Therefore, this paper contributes to the analytics literature by illustrating how date/time variables - in applications where the response variable is defined as the occurrence of some event within a certain period - can be used in data management to improve the performance of predictive models and the resulting decision support systems.