WJ-CH10-模式识别-聚类算法-04 英文版 讲解材料.ppt

Fuzzy Clustering – Hyperplanes as representatives Algorithms that recover hyperplanar clusters. Fuzzy c-varieties (FCV) algorithm It is based on the minimization of the distances of the vectors in X from hyperplanes. Disadvantage: It tends to recover very long clusters and, thus, collinear distinct clusters may be merged to a single one. Gustafson-Kessel (GK) algorithm Each planar cluster is represented by a center cj and a covariance matrix Σj, i.e., θj=(cj, Σj). The distance between a point x and the j-th cluster is defined as dGK2(x,θj)=|Σj|1/l(x-cj)TΣj-1(x-cj) The GK algorithm is derived via the minimization of the cost function Fuzzy Clustering – Hyperplanes as representatives (cont) The GK algorithm stems from GFAS. Setting the derivative of JGK(θ,U) with respect to cj’s and Aj’s equal to zero, the “parameter updating” part of the algorithm becomes: Example 5: In the first case, the clusters are well separated and the GK-algorithm recovers them correctly. In the second case, the clusters are not well separated and the GK-algorithm fails to recover them correctly. Possibilistic Clustering Unlike fuzzy clustering, the constraints on uij’s are uij ? [0, 1] maxj=1,…,m uij 0, i=1,…,N Possibilistic clustering algorithms result from the optimization of cost functions like where ηj are suitably chosen positive constants (see below). The 2nd term is inserted in order to avoid the trivial zero solution for the uij’ s. (other choices for the second term of the cost function are also possible (see below)). Setting ?J(θ,U)/ ?uij=0 we obtain: Possibilistic clustering (cont) Generalized Possibilistic Algorithmic Scheme (GPAS) Fix ηj, j=1,…,m. Choose θj(0) as the initial estimates of θj , j=1,…,m. t=0 Repeat For i=1 to N For j=1 to m End {For-j} End {For-i} t=t+1 Possibilistic clustering (cont) Generalized Possibilistic Algorithmic Scheme (GPAS) (cont) For j=1 to m Parameter updating: Solve with respect to θj and se