Learning from Mixture of Experimental Data: A Constraint–Based Approach

Vincenzo Lagani¹, Ioannis Tsamardinos^{1, 2}, and Sofia Triantafillou^{1, 2}

¹BioInformatics Laboratory - FORTH-ICS, Vassilika Vouton 100, Heraklion, Greece

²Computer Science Department, University of Crete, Knossou Ave., Heraklion, Greece

Abstract. We propose a novel approach for learning graphical models when data coming from different experimental conditions are available. We argue that classical constraint–based algorithms can be easily applied to mixture of experimental data given an appropriate conditional independence test. We show that, when perfect statistical inference are assumed, a sound conditional independence test for mixtures of experimental data can consist in evaluating the null hypothesis of conditional independence separately for each experimental condition. We successively indicate how this test can be modified in order to take in account statistical errors. Finally, we provide “Proof-of-Concept” results for demonstrating the validity of our claims.

Keywords: Graphical Models, Mixture of Experimental data, Conditional independence test, Constraint Based learning

LNAI 7297, p. 124 ff.

Full article in PDF | BibTeX

[email protected]
© Springer-Verlag Berlin Heidelberg 2012