Zhang and Cheung's new distance metric improves categorical clustering
Researchers differentiate nominal vs ordinal attributes for better distance measurement.
The paper addresses a fundamental gap in categorical data clustering: most methods treat nominal and ordinal attributes identically when computing dissimilarity, ignoring the relative order information in ordinal values. Zhang and Cheung introduce a graph-inspired perspective to model the intrinsic differences and connections between these attribute types. They propose a unified distance metric that measures intra-attribute distances for both nominal and ordinal values while preserving ordinal relationships.
Building on this metric, the authors develop a new clustering algorithm that jointly learns the intra-attribute distance weights and partitions data objects in a single optimization paradigm, rather than two separate steps. This integration circumvents suboptimal solutions common in sequential approaches. Evaluated on multiple real-world datasets, the proposed method demonstrates significant improvements in clustering accuracy and stability over existing categorical clustering algorithms like k-modes and its variants. The work is published in IEEE Transactions on Pattern Analysis and Machine Intelligence.
- Distinguishes nominal and ordinal categorical attributes for distance computation, preserving order information.
- Jointly learns intra-attribute distance weights and data partitions in a single optimization framework.
- Outperforms existing categorical clustering methods on benchmark datasets in terms of clustering accuracy.
Why It Matters
Improves clustering accuracy for real-world categorical data with mixed types, enabling better insights in analytics and machine learning pipelines.