Selection of variables structured by regularization in a multi-task framework

Sélection de variables structurée par régularisation jointe dans un cadre multi-tâches

Antoine Bonnefoy (1) , Ismael Ouamlil (2) , Jean-Baptiste Veyrieras (2) , Pierre Mahé (2)

1 QARMA - éQuipe AppRentissage et MultimediA [Marseille]
2 Bio-Mérieux [Marcy l'Etoile]

Motivated by diagnostic applications in the field of clinical microbiology, we introduce a joint in-put/output regularization method to perform struc-tured variable selection in a multi-task setting where tasks can exhibit various degrees of correlation. Our approach extensively relies on the tree-structured group-lasso penalty and explicitly combines hierarchical structures defined across features and task by means of the Cartesian product of graphs to induce a global hierarchical group structure. A vectorization procedure is then used to solve the resulting multi-task problem with standard mono-task optimization algorithms developed for the overlapping group-lasso problem. Experimental results on simulated and real data demonstrate the interest of the approach.

Mots clés: clinical microbiology; hierarchical; optimization algorithms

Domaines: Traitement du signal et de l'image [eess.SP]; Apprentissage [cs.LG]; Bio-informatique [q-bio.QM]

Antoine Bonnefoy, Ismael Ouamlil, Jean-Baptiste Veyrieras, Pierre Mahé. Selection of variables structured by regularization in a multi-task framework. Conférence francophone sur l’apprentissage automatique (CAp), Jul 2016, Marseille, France. ⟨hal-01373190⟩