Representative noise-free complete-link classification with application to protein structures

Abstract

In various applications, including many problems of knowledge discovery in databases, and particularly in the field of computational molecular biology, a compact and representative description of a vast object space is desired. In this paper, a constructive mathematical model corresponding the intuitive requirements of representativity is developed. Representativity is divided into two aspects: typicality and comprehensiveness. A new sieving method is presented where a special kind of noise is detected and eliminated by removing anomalous objects from the initial complete linkage partition. The comprehensiveness endangered by sieving is then regained by applying a special completion procedure. Theoretical results ensure that the resulting partition is representative, consisting of solid and separable classes. The conceptual model was further tested by applying the method to protein amino acid sequences of the Brookhaven Protein Data Bank. The recognized biochemical substance of the outcome confirm the representativity of the resulting classification.