Integrated Learning via Randomized Forests and Localized Regression with Application to Medical Diagnosis

The tree-based machine learning functions on the divide-and-conquer principle and is known to perform well in certain applications. In this paper, we first give a new data partitioning rule using the mean of the data columns to grow the tree till the child nodes are small in size. Then, the local regression is applied to leave nodes to enhance the resolution of the node outputs. Randomization is introduced at tree growth and forest creation. The local prediction accuracies on the leaves are used to select a subset of the test data for actual predictions. The case study on the diagnosis of autistic spectrum disorder shows that the proposed method achieves the prediction accuracy of the ensemble at above 96% with reduced variance, which is much better than those reported in the literature.