An Experimental Study of K* Algorithm

Abstract

Machine Learning techniques are taking place in all areas of our lives, to help us to make decisions. There is a large number of algorithms available for multiple purposes and appropriate for specific data types. That is why it is required to pay special attention to decide which is the recommended technique, to use in each case. K Star is an instance-based learner that tries to improve its performance for dealing with missing values, smoothness problems and both real and symbolic valued attributes; but it is not known much information about how the way it faces attribute and class noisy, and with mixed values of the attributes in the datasets. In this paper we made six experiments with Weka, to compare K Star and other important algorithms: Nave Bayes, C4.5, Support Vector Machines and k-Nearest Neighbors, taking into account its performance classifying datasets with those features. As a result, K Star demonstrated to be the best of them in dealing with noisy attributes and with imbalanced attributes.