Dual Weighted Inverted Specific-Class Distance Measure for Nominal Attributes

Abstract

The inverted specific-class distance measure (ISCDM) continues to be one of the top distance metrics addressing only address nominal attributes when missing values and noises exist in the training set. However, its attribute independence assumption fails to hold true in applications with sophisticated attribute dependencies. To alleviate this assumption, several attribute weighting-based and instance weighting-based improved versions have been proposed. However, all of them are limited to focusing solely on either attribute weighting or instance weighting; they ignore the enhancement of the measuring performance by the combination of attribute weighting and instance weighting. Thus, in this study, we developed a novel dual-weighting scheme that considers attribute weighting and instance weighting simultaneously, denoted as dual weighted ISCDM (DWISCDM). We incorporate the attribute and instance weights into the difference formulas of two attribute values and instances, respectively. Extensive experimental results demonstrate the effectiveness of the proposed DWISCDM against ISCDM and some existing state-of-the-art baselines in terms of negative conditional log likelihood and root relative squared error.

Keywords: Distance metric learning, inverted specific-class distance measure, dual weighting, attribute weighting, instance weighting