Aggregation of Random Initializations for Robust Linear Clustering

Outliers Álvarez-Esteban, P.C. García-Escudero, L.A. Mayo-Iscar, A. can be very harmful to clustering methods, to the extent that even a small fraction of outliers can significantly impact their performance. Consequently, the use of robust clustering procedures has been advocated. A robust linear clustering procedure will be reviewed here, where robustness is pursued by applying a trimming approach. The proposed algorithm for its practical implementation is based on concentration steps, analogous to those applied in well-known high-breakdown robust multivariate procedures. However, correct initialization is key for the algorithm’s adequate performance. A large number of random initializations is often required to guarantee avoidance of local minima of the robust linear clustering target function. An alternative “ensemble initialization” method will be given to try to overcome this problem.