J. E. Chacón
The mean shift algorithm was introduced in Fukunaga and Hostetler (1975). It is an iterative procedure which, at every step, shifts the point obtained in the previous iteration in the direction of the estimated normalized density gradient, producing a convergent sequence that transports the initial value to the closest local maximum of the density estimate along the steepest ascent path. This algorithm induces a partition of the data in a natural way, by assigning the same cluster to all the data points that lead to the same local maximum when the convergence of the iterative procedure is reached. Notice that this methodology does not require the number of clusters to be specified in advance, and that it allows clusters of arbitrary shape to be discovered. In this communication we study the properties of this algorithm when kernel methods are used to estimate the normalized density gradient, and propose several data-driven bandwidth selectors to make this methodology fully automatic.
Palabras clave: bandwidth selection, mean shift clustering, kernel methods, density gradient estimation
Programado
VA7 Estadística no paramétrica
20 de abril de 2012 09:00
Sala Roma II
