Funded Projects

Bayesian Matrix and Tensor factorisation methods

TUBITAK 110E292: Fast Bayesian Matrix and Tensor factorisation methods for nonstationary multivariate time series analysis (BAYTEN)
Dates: 2 Years, April 2011-April 2013

In this project, we develop novel computational methodology for modeling and inference in multivariate and multidimensional time series data. Our final goal is developing effective and efficient computational tools that scale well with data, that are useful in a wide spectrum of application areas such as statistical signal processing, machine learning, information retrieval, bioinformatics, sensor networks, seismic data analysis, surveillance and network analysis.
The scientific project is focused on matrix and tensor factorization techniques in statistical settings and with the help of a graphical modelling formalism. The statistical framework allows to make the data modelling assumptions explicit. The graphical formalism will make the methodology fairly general, flexible and most importantly accessible. Our approach will support and will be supported by the development of a parallel Bayesian Tensor factorisation library, that will facilitate Bayesian approaches allowing for regularization of the matrix factors and model order selection, as well as to computational Bayes methods such as MCMC and particle filtering. In particular the matrix/tensor factorization problem brings out novel methodological problems related to the problem of inference in composite models that will be important in many applied and computational areas.