Checking Synthetic Instances By Own for Better Classification in Decision Tree

摘要

Because decision tree algorithms give the priority to the classes having more training instances for better classification accuracy, supplying more training instances for a specific class may improve the classification accuracy of the class. Synthetic minority over-sampling technique supplies instances of a minor or rare class to build better classification models for the minor class. But, if we build a decision tree using a training data set, some instances are classified wrongly. There are two reasons for the wrong classification - the limitation of the data mining algorithm itself, and imperfection of the data set. As a way to build better decision tree for a minority class without sacrificing overall accuracy much, we select good synthetic data instances for our decision tree. By checking whether the synthetic data instances are classified correctly, and supplying the good ones only to build our target decision tree, we could build better decision tree for a minor class. Experiments were done using a data set in the domain of liver disease proved our assertion.