On the Abstraction of a Categorical Clustering Algorithm

Despite being one of the most common approach in unsupervised data analysis, a very small literature exists on the formalization of clustering algorithms. This paper proposes a semiring-based methodology, named Feature-Cluster Algebra, which is applied to abstract the representation of a labeled tree structure representing a hierarchical categorical clustering algorithm, named CCTree. The elements of the feature-cluster algebra are called terms. We prove that a specific kind of a term, under some conditions, fully abstracts a labeled tree structure. The abstraction methodology maps the original problem to a new representation by removing unwanted details, which makes it simpler to handle. Moreover, we present a set of relations and functions on the algebraic structure to shape the requirements of a term to represent a CCTree structure. The proposed formal approach can be generalized to other categorical clustering (classification) algorithms in which features play key roles in specifying the clusters (classes).