Timezone: »
We consider the problem of binary classification where the classifier may abstain instead of classifying each observation. The Bayes decision rule for this setup, known as Chow's rule, is defined by two thresholds on posterior probabilities. From simple desiderata, namely the consistency and the sparsity of the classifier, we derive the double hinge loss function that focuses on estimating conditional probabilities only in the vicinity of the threshold points of the optimal decision rule. We show that, for suitable kernel machines, our approach is universally consistent. We cast the problem of minimizing the double hinge loss as a quadratic program akin to the standard SVM optimization problem and propose an active set method to solve it efficiently. We finally provide preliminary experimental results illustrating the interest of our constructive approach to devising loss functions.
Author Information
Yves Grandvalet (Université de Technologie de Compiégne)
Joseph Keshet (Bar-Ilan University)
Alain Rakotomamonjy (Université de Rouen Normandie Criteo AI Lab)
Stephane Canu (INSA Rouen, LITIS)
More from the Same Authors
-
2019 Poster: Screening Sinkhorn Algorithm for Regularized Optimal Transport »
Mokhtar Z. Alaya · Maxime Berar · Gilles Gasso · Alain Rakotomamonjy -
2019 Poster: Singleshot : a scalable Tucker tensor decomposition »
Abraham Traore · Maxime Berar · Alain Rakotomamonjy -
2017 Poster: Joint distribution optimal transportation for domain adaptation »
Nicolas Courty · Rémi Flamary · Amaury Habrard · Alain Rakotomamonjy -
2014 Poster: Optimizing F-Measures by Cost-Sensitive Classification »
Shameem Puthiya Parambath · Nicolas Usunier · Yves Grandvalet -
2013 Poster: Learning Efficient Random Maximum A-Posteriori Predictors with Non-Decomposable Loss Functions »
Tamir Hazan · Subhransu Maji · Joseph Keshet · Tommi Jaakkola -
2012 Poster: Multiple Operator-valued Kernel Learning »
Hachem Kadri · Alain Rakotomamonjy · Francis Bach · philippe preux -
2011 Poster: Generalization Bounds and Consistency for Latent Structural Probit and Ramp Loss »
David Mcallester · Joseph Keshet -
2011 Oral: Generalization Bounds and Consistency for Latent Structural Probit and Ramp Loss »
David Mcallester · Joseph Keshet -
2010 Workshop: New Directions in Multiple Kernel Learning »
Marius Kloft · Ulrich Rueckert · Cheng Soon Ong · Alain Rakotomamonjy · Soeren Sonnenburg · Francis Bach -
2010 Poster: Direct Loss Minimization for Structured Prediction »
David A McAllester · Tamir Hazan · Joseph Keshet -
2009 Workshop: Temporal Segmentation: Perspectives from Statistics, Machine Learning, and Signal Processing »
Stephane Canu · Olivier Cappé · Arthur Gretton · Zaid Harchaoui · Alain Rakotomamonjy · Jean-Philippe Vert -
2007 Spotlight: Hierarchical Penalization »
Marie Szafranski · Yves Grandvalet · Pierre Morizet-Mahoudeaux -
2007 Poster: Hierarchical Penalization »
Marie Szafranski · Yves Grandvalet · Pierre Morizet-Mahoudeaux