Associative Clustering

This report contains derivations which did not fit into the paper [3]. Associative clustering (AC) is a method for separately clustering two data sets when one-to-one associations between the sets, implying statistical dependency, are available. AC finds Voronoi partitionings that maximize the visibility of the dependency on the cluster level. The main content of this paper are technical results related to the algorithm: A Bayes factor interpretation of AC, derivation of gradients for optimizing AC with a smoothing trick, and the connection of AC objective to mutual information.

• Type: Conference
• Year: 2004
• Where: ECML
• Authors: Janne Sinkkonen, Janne Nikkilä, Leo Lahti, Samuel Kaski
• Keywords: AC Objective, Associative Clustering, ECML 2004, Report Contains Derivations