Join us for an invited talk with Bernhard Geiger (TU Munich, Germany)

• Title: “Information-Theory for Markov Aggregation and Clustering”
• Event Date: September 6, 12:00-13:00.
• Location: Room 4.2.E03; Torres Quevedo Building; Leganés Campus; Universidad Carlos III de Madrid.

Abstract:

In many scientific disciplines, Markov models are too large to allow efficient simulation or parameter estimation – for example, in natural language processing and chemical reaction networks, the number of states in such models is linked to the number of n-tuples of words in a dictionary and to the number of molecules in a given volume, respectively. A possible way to cope with this problem is aggregation, i.e., building a smaller Markov model by clustering states, thus trading model complexity for model accuracy.
We present two different information-theoretic cost functions for Markov aggregation, which are linked to predictability and lumpability, the phenomenon in which a function of a Markov chain has the Markov property. We then show that the Markov aggregation problem generalizes several information-theoretic clustering problems, such as the information-theoretic co-clustering by Dhillon.