THE USAGE OF METRIC FEATURES IN PREDICTION WITH DECISION TREES DEMONSTRATED ON THE TASK OF FOREST COVER TYPE CLASSIFICATION

Abstract

Methods of classification by nature of decision-making divide on methods using global optimization (all training samples are used), and local optimization (only samples in the neighbourhood of the studied object are used). The perspective direction of research is combination of advantages of each approach in one integrated classifier. In article the method of combination of these approaches by embedding of local metric features into the approach using global optimization is proposed. This approach is shown for a case when the classifier using global optimization is random forest and extra random trees. Various variants of metric features are evaluated. Performance of the proposed approach is illustrated on the forest cover type prediction task, where it leads to significant improvement in classification accuracy.