Unsupervised object ranking using not even weak experts

Antoine Cornuéjols (1) , Christine Martin (1)

1 MIA-Paris - Mathématiques et Informatique Appliquées

Many problems, like feature selection, involve evaluating objects while ignoring the relevant underlying properties that determine their true value. Generally, an heuristic evaluating device (e.g. filter, wrapper, etc) is then used with no guarantee on the result. We show in this paper how a set of experts (evaluation function of the objects), not even necessarily weakly positively correlated with the unknown ideal expert, can be used to dramatically improve the accuracy of the selection of positive objects, or of the resulting ranking. Experimental results obtained on both synthetic and real data confirm the validity of the approach. General lessons and possible extensions are discussed.

Mots clés: ensemble methods; unsupervised learning; feature selection; ranking; apprentissage libre

Domaines: Sciences du Vivant [q-bio]

Antoine Cornuéjols, Christine Martin. Unsupervised object ranking using not even weak experts. Neural Information Processing 2011: Proceedings, 18th International Conference, ICONIP 2011 Shanghai, China, November 13-17, 2011 Proceedings, Part I, 7062, Springer - Verlag, 2011, Lecture Notes in Computer Science, 978-3-642-24954-9. ⟨10.1007/978-3-642-24955-6_72⟩. ⟨hal-01197561⟩