Hybridizing Extremely Randomized Trees with Bootstrap Aggregation for Crime Prediction

The prevalence of crime continues to be a major challenge in communities and societies around the globe. This justifies the relevance of studies on crime prevention. As a preventive strategy, crime prediction can help deter known crimes before they occur. Machine learning algorithms have been vastly applied to predictive tasks, particularly Decision Trees (DT), among others. Despite their good performance, DT suffers from bias and variance problems. While DT has these problems, there are other algorithms, which are variants of DT that are more viable. The two algorithms known to reduce bias and variance are Random Forest (RF) and Extremely Randomized Trees (ERT). In this work, we proposed a hybrid algorithm which utilizes the best attributes from both RF and ERT, which are bootstrap aggregation and random features selection. We then compared our hybrid algorithm with RF and ERT. Obtained results show that our hybrid algorithm performed better in terms of prediction accuracy, and computational complexity.