An Improved Nominal Attribute Distance Measure Based on Attribute Dependent Representation

In real data, there are dependencies between at-tributes, and most distance metrics make attribute independence assumptions, and although they show superior performance, when there are strong dependencies between attributes, it causes a certain degree of degradation to the distance metric performance. For this reason, this paper proposes a nominal attribute distance metric based on attribute dependencies. We conduct an extensive evaluation on a large number of UCI datasets and find that the method significantly outperforms currently available structural extension schemes in terms of negative log-likelihood and root-relative-square, while ensuring the robustness of the algorithm.