Approximation of Data by Decomposable Belief Models

It is well known that among all probabilistic graphical Markov models the class of decomposable models is the most advantageous in the sense that the respective distributions can be expressed with the help of their marginals and that the most efficient computational procedures are designed for their processing (for example professional software does not perform computations with Bayesian networks but with decomposable models into which the original Bayesian network is transformed). This paper introduces a definition of the counterpart of these models within Dempster-Shafer theory of evidence, makes a survey of their most important properties and illustrates their efficiency on the problem of approximation of a “sample distribution” for a data file with missing values.

• Type: Journal
• Year: 2010
• Where: IPMU
• Authors: Radim Jirousek
• Keywords: Bayesian Network, Decomposable Models, Graphical Markov Models, Information Technology, IPMU 2010