Abstract
The success of categorical data clustering generally relies heavily on the distance metric that measures the dissimilarity between two objects. However, most existing clustering methods treat the two categorical subtypes, nominal and ordinal attributes, similarly when calculating dissimilarity, neglecting the relative order information of ordinal values. Furthermore, interdependence among nominal and ordinal attributes is worth exploring to indicate dissimilarity.
This paper studies the intrinsic differences and connections of nominal and ordinal attribute values from a graph-like perspective. Accordingly, we propose a novel distance metric to measure the intra-attribute distances of nominal and ordinal attributes in a unified manner while preserving the order relationship among ordinal values. Subsequently, we introduce a new clustering algorithm that integrates the learning of intra-attribute distance weights and data object partitions into a single learning paradigm, circumventing a suboptimal solution.
Experiments demonstrate the efficacy of the proposed algorithm compared to existing counterparts.
Blogger's Review: This paper innovatively addresses the limitations of traditional methods by integrating distance metrics for nominal and ordinal attributes. This approach not only enhances clustering accuracy but also provides new insights for future research in this area, making it a noteworthy contribution.