Gain ratio weighted inverted specific-class distance measure for nominal attributes

Abstract

Enhancing distance measures is key to improving the performances of many machine learning algorithms, such as instance-based learning algorithms. Although the inverted specific-class distance measure (ISCDM) is among the top performing distance measures addressing nominal attributes with the presence of missing values and non-class attribute noise in the training set, this still requires the attribute independence assumption. It is obvious that the attribute independence assumption required by the ISCDM is rarely true in reality, which harms its performance in applications with complex attribute dependencies. Thus, in this study we propose an improved ISCDM by utilizing attribute weighting to circumvent the attribute independence assumption. In our improved ISCDM, we simply define the weight of each attribute as its gain ratio. Thus, we denote our improved ISCDM as the gain ratio weighted ISCDM (GRWISCDM for short). We tested the GRWISCDM experimentally on 29 University of California at Irvine datasets, and found that it significantly outperforms the original ISCDM and some other state-of-the-art competitors in terms of the negative conditional log likelihood and root relative squared error.