Conditional Information Bottleneck Clustering

We present an extension of the well-known information bottleneck framework, called conditional information bot- tleneck, which takes negative relevance information into account by maximizing a conditional mutual information score. This general approach can be utilized in a data mining context to extract relevant information that is at the same time novel relative to known properties or structures of the data. We present possible applications of the condi- tional information bottleneck in information retrieval and text mining for recovering non-redundant clustering solu- tions, including experimental results on the WebKB data set which validate the approach.