Sequential Information Bottleneck for Finite Data
Jaakko Peltonen - Helsinki University of Technology, Neural Networks Research Centre Janne Sinkkonen - Helsinki University of Technology, Neural Networks Research Centre Samuel Kaski - Helsinki University of Technology, Neural Networks Research Centre
The sequential information bottleneck (sIB) algorithm clusters co-occurrencedata such as text documents vs. words. We introduce a variant that modelssparse co-occurrence data by a generative process. This turns the objectivefunction of sIB, mutual information, into a Bayes factor, while keeping itintact asymptotically, for non-sparse data. Experimental performance of thenew algorithm is comparable to the original sIB for large data sets, andbetter for smaller, sparse sets.