Robust clustering methods based on the tau scale: preliminary results.

GONZALEZ, JUAN DOMINGO; YOHAI, VICTOR J.

Buenos Aires

Encuentro; VI Escuela de Invierno Luis A. Santaló; 2014

Instituto de Cálculo

Cluster Analysis methods aim to detect homogeneous groups within a population. It is well known that a single outlier in the data may have a large effect on the selected clusters when they are obtained using a traditional procedure as, for example, the K-Means algorithm. To overcome this problem, we propose a clustering procedure based on the minimization of an efficient and robust scale (the tau scale proposed by Yohai and Zamar (1988)) of the distances between the points and the corresponding cluster center. To evaluate the proposed clustering algorithm, numerical experiments have been carried using several data sets. The results show that the proposed procedure based on the minimization of the tau-scale compares favorably with respect to other widely used algorithms as K-Means and Trimmed-K-Means. This is a joint work in progress with Victor Yohai and Ruben Zamar.