Characterizations of Admissible Objective Functions for Hierarchical Clustering

TL;DR

This paper characterizes admissible objective functions for hierarchical clustering, focusing on sum-type and max-type classes, and analyzes their approximation guarantees with recursive algorithms.

Contribution

It provides a complete characterization of admissible sum-type functions with polynomial scaling and introduces max-type functions with their admissibility criteria.

Findings

Characterized admissible sum-type functions with polynomial g of degree ≤ 2.

Established approximation guarantees for recursive sparsest cut algorithms.

Provided a complete characterization of admissibility for max-type functions.

Abstract

Hierarchical clustering is a fundamental task in data analysis, yet for a long time it lacked a principled objective function. Dasgupta [STOC 2016] initiated a formal framework by introducing a discrete objective function for cluster trees. This framework was subsequently expanded by Cohen-Addad et al. [J. ACM 2019], who introduced the notion of admissibility -- a criterion ensuring that, whenever the input similarities admit consistent hierarchical representations, the minimizers of an objective function recover them. They also provided a necessary and sufficient condition for admissibility within a broad class of objective functions, which we refer to as sum-type objective functions. Our contributions are twofold. First, we characterize admissible sum-type objective functions when the scaling function g is a symmetric polynomial of degree at most two, together with sufficient conditions in the degree-three case. For admissible objective functions in this class, we show that the recursive sparsest cut algorithm achieves an O($\phi$)-approximation ratio, where $\phi$ denotes the approximation factor of the sparsest cut subroutine. Second, we introduce a new class of objective functions for hierarchical clustering, which we term max-type objective functions, where cluster interactions are measured by maximum rather than aggregate similarity. For this class, we establish a general characterization of admissibility for arbitrary scaling functions g, and a complete characterization when g is a symmetric polynomial of degree at most two. These results provide new theoretical insights into admissible objective functions for hierarchical clustering and clarify the scope of algorithmic guarantees for optimizing them.