Gaining knowledge of data structure using stability concepts

A.E. BAYÁ; P. M. GRANITTO

Mar del Plata, Argentina

Congreso; ARGENTINE SYMPOSIUM ON ARTIFICIAL INTELLIGENCE (ASAI); 2009

Sociedad Argentina de Informatica

In this work we propose the use of stability concepts to gain knowledge on data structure. To analyze stability we use an exploration algorithm based on the idea of ``Sub-sample and Cluster´´. In the clustering step we use PAM and HC-av to construct solution ensembles and then we evaluate their similarity with the Variation of Information Criterion. We apply our method to analyze artificial and real data. We use some datasets to compare our method with another analysis technique proposed by Ben-Hur and Guyon. We show that we can obtain comparable information with both methods but our new technique presents results that can be more easily analyzed.