Amplifying the Block Matrix Structure for Spectral Clustering

This page contains reference data and code used in a paper by Jan Poland and myself, published back in 2005.

Data Sets Used

This benchmark data for spectral clustering algorithms. are organized in Matlab .mat files, each containing the following variables:

x the matrix of the data themselves, as column vectors
c0 a row vector of correct cluster assignments, an int for each point
nn a row vector of cluster sizes, an int for each cluster
sgm a matrix of best kernel widths (sigma) for every algorithm. Each row contains five empirically found best values for the corresponding algorithm. The algorithms are ordered as follows:
  1. KEM
  2. Ng & al.
  3. Basic spectral
  4. Conductivity
  5. Laplacian
sgmHist a vector of most probable sigmas according to the histogram method

The data sets are:

Programs Used

The algorithms were implemented---with one exception for performance reasons---in Matlab. The programs files are:

Zip Download

You may download all the files as a single zip file.