Fine-grained attribute weighted inverted specific-class distance measure for nominal attributes

Abstract The inverted specific-class distance measure (ISCDM) ranks first in the list of distance metrics that deal solely with nominal attributes, especially when values are missing and noise exists in non-class attributes. However, the attribute independence assumption still inevitably exists, which is almost untenable in many real-world applications with sophisticated attribute dependencies. Many improved versions based on different attribute weighting schemes have been proposed to relax this unrealistic assumption and circumvent its damage to measuring performance. However, existing attribute weighting schemes are limited to assigning a weight corresponding to each attribute; they ignore the more fine-grained dependence relationships between attributes and classes. Thus, in this study, we derive a novel fine-grained attribute weighting scheme, which first calculates the initial fine-grained attribute weights according to different attribute values and class labels and then uses random walk with restart to optimize them. We titled our improved measure fine-grained attribute-weighted ISCDM (FAWISCDM). Extensive experimental results on 66 datasets from a machine learning repository, collected by the University of California at Irvine illustrate that the FAWISCDM is notably superior to the original ISCDM and some other state-of-the-art competing methods, in terms of the negative conditional log likelihood and root relative squared error.