A probabilistic approach to spectral clustering: Using KL divergence to find good clusters
Guido Sanguinetti, Jonathan Laidler and Neil Lawrence
In: Pascal Statistics and Optimization of Clustering Workshop 2005, 4-5 Jul 2005, London, UK.
In previous work we have demonstrated a heuristic approach to
spectral clustering which automatically determines the number of
clusters that appear in the data set. Here we present a possible
probabilistic interpretation for this, where the correct
clustering can be found by minimising the Kullback-Leibler (KL)
divergence between the affinity matrix and an appropriate Gaussian approximation.