An automatized approach to obtaining input parameters in robust model based clustering

A. Mayo-Iscar, L. Á. García-Escudero, C. Hennig, G. Morelli, M. Riani

TCLUST constitutes a robust model based clustering methodology developed over the last 15 years, where robustness is based on the joint application of trimming and constraints. A neverending open issue when applying model based clustering is linked to the number of groups (k). In relation to TCLUST application you should take into account additional input parameters linked to the proportion of trimming (alpha) and to the level of the constraints (c). Now we are proposing a bootstrap method designed to generate a reduced list of sensible pairs of cadidates for the k and alpha parameters. This automated approach, which tries to replace a current graphical tool, eliminates the need for visual inspection of the curves and minimizes subjective decisions associated with sample variability. After making the mentioned reduced list available, users can select the robust cluster partition that best aligns with their specific clustering objectives by utilizing standard cluster validation tools.

Keywords: clustering, constraints, robustness, trimming

GT03.AMC4 Clustering and Classification
November 9, 2023  3:30 PM
CC3: Room 1

D. Rodríguez Vítores, C. Matrán Bea

D. Peña

E. Boj del Val, A. Grané