Dirichlet Mixtures of Bayesian Linear Gaussian State-Space Models


We give a MATLAB implementation of the Dirichlet Mixtures of Bayesian Linear Gaussian State-Space Models (DMBLGSSM) using a variational approach. This model can be used to perform unsupervised clustering of multidimansional time-series.The model uses priors on the LGSSM to eliminate unnecessary parameters and a prior on the mixture components to determine an appropriate number of clusters.

We consider two variants of this model which can be more or less appropriate depending on the application. In the first variant, time-series are clustered together when they show globally similar dynamics. In the second variant, time-series are clustered together when they show simultaneously similar dynamics.  In other words, in the fisrt case time-series are clustered together when they are generated by a different realization of the same dynamical process, while in the second case time-series are clustered together when they are generated by the same realization of a dynamical process.


DIRICHLET MIXTURES OF BAYESIAN LGSSMS: TIME-SERIES CLUSTERING BASED ON GLOBAL SIMILARITY 


DMBLGSSM  for time-series clustering based on global similarity (DMBLGSSM.zip ) (Last updated: 28 November 2007).

This code can deal with missing observations (DMBLGSSM_MISS.zip ) (Last updated: 28 November 2007).


DIRICHLET MIXTURES OF BAYESIAN LGSSMS:  TIME-SERIES CLUSTERING BASED ON SIMULTANEOUS SIMILARITY 


DMBLGSSM for time-series clustering based on simultaneous similarity (DMBLGSSMO.zip ) (Last update: 28 November 2007).


References


Dirichlet Mixtures of Bayesian Linear Gaussian State-Space Models: a Variational Approach. S. Chiappa and D. Barber. Technical report no. 161, Max-Planck Institute for Biological Cybernetics,Tübingen, Germany, 2007 [ .pdf ]

 

Output Grouping using Dirichlet Mixtures of Linear Gaussian State-Space Models . S. Chiappa and D. Barber, ISPA, 2007 [ .pdf ]

Make a Free Website with Yola.