Improving Generalization Ability of Classifier with Multiple Imputation Techniques

Abstract

A main objective of research in machine learning is to learn to identify complex patterns and make intelligent decisions based on that data automatically; but the set of all possible behaviors is too large to be covered by the set of available training data. Hence it is desirable that learner must generalize from the given examples, so that it can provide a useful output in new cases; otherwise the number of training instances available in training data must be sufficient. We can have another case where the data set may contain some instances with multiple missing attributes, these instances need to be deleted; in such case sufficient data samples are required to improve generalization ability of the classifier. The proposed algorithm generates additional training instances and adds it to original training data to improve generalization ability of the decision tree classifiers. The proposed algorithm imputes missing attribute values with domain values and thus generates additional training instances. The proposed method is permutation and combination based multiple imputation method and it is also useful for imputation of missing data. The proposed method demonstrates good generalization ability on decision trees. This paper proposes a new method for imputation of missing data and same method is used to generate additional data instances to generalize the decision tree learning.