The Adaptive Mean-Linkage Algorithm: A Bottom-Up Hierarchical Cluster Technique

TL;DR

The paper introduces the Adaptive Mean-Linkage Algorithm, a bottom-up hierarchical clustering method that dynamically updates thresholds to improve cluster homogeneity and merging efficiency, demonstrated on chemical substituents.

Contribution

It presents a novel adaptive threshold updating strategy for hierarchical clustering, enhancing performance over traditional average-linkage methods.

Findings

Achieves better cluster homogeneity.

Ensures at least one pair of sets merges at each step.

Demonstrated effectiveness on chemical data.

Abstract

In this paper a variant of the classical hierarchical cluster analysis is reported. This agglomerative (bottom-up) cluster technique is referred to as the Adaptive Mean-Linkage Algorithm. It can be interpreted as a linkage algorithm where the value of the threshold is conveniently up-dated at each interaction. The superiority of the adaptive clustering with respect to the average-linkage algorithm follows because it achieves a good compromise on threshold values: Thresholds based on the cut-off distance are sufficiently small to assure the homogeneity and also large enough to guarantee at least a pair of merging sets. This approach is applied to a set of possible substituents in a chemical series.