June, Thursday 28th
14:30 (Shannon amphitheatre, 660 building)
(LIRIS - Lyon)
Title: Identifying irreducible disjoint factors in multivariate probability distributions: Application to multilabel learning
In this talk, I discuss the problem of decomposing a multivariate probability distribution into a product of factors defined over disjoint subsets of random variables called irreducible disjoint factors (IDFs). I show that the IDFs are the connected components of an undirected graphical model which structure can be inferred by running a series of conditional independence tests. Several theoretical results are established to characterize the IDFs under various assumptions about the probability distribution (i.e., DAG-Faithfulness, Intersection property, Composition property). I show how the IDF decomposition can help to identify the bayes optimal solution of the multi-label classification problem using the subset zero-one loss or the F-measure.
Contact: guillaume.charpiat at inria.fr
All TAU seminars: here