Optimism in Active Learning

Abstract EN

Active learning is the problem of interactivelyconstructing the training set used in classificationin order to reduce its size.

It would ideallysuccessively add the instance-label pairthat decreases the classification error most.

However,the effect of the addition of a pair is notknown in advance.

It can still be estimatedwith the pairs already in the training set.

Theonline minimization of the classification errorinvolves a tradeoff between exploration andexploitation.

This is a common problem inmachine learning for which multiarmed bandit,using the approach of Optimism int the Face of Uncertainty, has proven very efficient these lastyears.

This paper introduces three algorithmsfor the active learning problem in classificationusing Optimism in the Face of Uncertainty.

Experiments lead on built-in problems and realworld datasets demonstrate that they comparepositively to state-of-the-art methods.