Deconvolutive clustering of markov states

Authors:
Ata Kabán;Xin Wang
Affiliations:
School of Computer Science, The University of Birmingham, Birmingham, UK;School of Computer Science, The University of Birmingham, Birmingham, UK
Venue:
ECML'06 Proceedings of the 17th European conference on Machine Learning
Year:
2006

Citing 8
Cited 2

An information-maximization approach to blind separation and blind deconvolution

Neural Computation
Blind source separation and deconvolution: the dynamic component analysis algorithm

Neural Computation
Mixed Memory Markov Models: Decomposing Complex Stochastic Processes as Mixtures of Simpler Ones

Machine Learning
Pattern Recognition and Neural Networks

Pattern Recognition and Neural Networks
Topic Identification in Dynamical Text by Complexity Pursuit

Neural Processing Letters
Learning to Probabilistically Identify Authoritative Documents

ICML '00 Proceedings of the Seventeenth International Conference on Machine Learning
Communities of Interest

IDA '01 Proceedings of the 4th International Conference on Advances in Intelligent Data Analysis
Uniformization and hypergraph partitioning for the distributed computation of response time densities in very large Markov models

Journal of Parallel and Distributed Computing

Factorisation and denoising of 0-1 data: A variational approach

Neurocomputing
State aggregation in higher order markov chains for finding online communities

IDEAL'06 Proceedings of the 7th international conference on Intelligent Data Engineering and Automated Learning

Quantified Score

Hi-index	0.01

Visualization

Abstract

In this paper we formulate the problem of grouping the states of a discrete Markov chain of arbitrary order simultaneously with deconvolving its transition probabilities. As the name indicates, this problem is related to deconvolutive blind signal separation. However, whilst the latter has been studied in the context of continuous signal processing, e.g. as a model of a real-room mixing of sound signals, our technique tries to model computer-mediated group-discussion participation from a discrete event-log sequence. In this context, convolution occurs due to various time-delay factors, such as the network transmission bandwidth or simply the typing speed of the participants. We derive a computationally efficient maximum likelihood estimation algorithm associated with our model, which exploits the sparsity of state transitions and scales linearly with the number of observed higher order transition patterns. Results obtained on a full day worth dynamic real-world Internet Relay Chat participation sequence demonstrate the advantages of our approach over state grouping alone, both in terms of penalised data likelihood and cluster clarity. Other potential applications of our model, viewed as a novel compact approximation of large Markov chains, are also discussed.