spclust

Spectral Clustering algorithm by Ng, Jordan and West (NIPS 2001)
[IDX,SUMD,DEG] = SPCLUST(X,K,VARARGIN)

 Returns:
    (IDX): Cluster assignment vector
    (SUMD): Within-cluster sums of point-to-centroid distances after the application
        of k-means to the K eigenvectors
    (U):   Eigenvectors of normalised Laplacian
    (DEG): Vector containing degree of each vertex


 Inputs:
    (X): N-by-D Data matrix or N-by-N Distance Matrix, 
        or 1-by-(N choose 2) vector of pairwise distances produced by pdist(DataMatrix).
        By default X is assumed to be a data matrix    
    (K): Number of clusters
    Optional Input Arguments specified as Name,Value pairs:
    (s): Scaling parameter used in similarity matrix:
            A_{ij} = exp( -norm(X(:,i) - X(:,j))^2/(2*s)^2 )
        If s=0 (which is the default) the local bandwidth selection rule of 
        Zelnik-Manor and Perona (NIPS 2004) is used
    (distmat): If TRUE then X is not a data matrix. Depending on its dimensionality X is treated as either 
        the output of pdist(DataMatrix) (1-by-(N choose 2)) or as a distance matrix (N-by-N)
        By default: distmat=FALSE
    (normalise): If TRUE and if X is a data matrix then all variables are scaled in [-1,1]
        By default normalise=FALSE
    (nn): Number of Nearest Neighbours to use in the computation of the local bandwidth selection rule
        Following Zelnik-Manor and Perona (NIPS 2004) the default value is 7.

References:
A.Y. Ng, M.I. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm.
Advances in Neural Information Processing Systems 14, pages 849-856. 2001.

L. Zelnik-Manor and P. Perona. Self-tuning spectral clustering.
Advances in Neural Information Processing Systems, pages 1601--1608, 2004.