Extreme value theory in open-set classification

Details

Supervisors

Segers, Johan

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

Statisticians, data scientists, and other data explorers all used to base and build their models on existing data. But what would happen if the focus was to handle unseen, or unknown observations ? This process is known as open-set classification in the context of statistical classification, or more generally as novelty detection when applied to more diverse research fields. The novelty detection, or open-set classification encompasses a series of techniques all aimed at classifying a test data-set as part of the known or unknown class. This is done by building a model from a training data-set considering all training instances as known observations. This approach is typically used when instances of the new or unknown class are in short supply. This master thesis will start by introducing the theory behind the statistical concepts of extreme value theory and classification. A review of existing novelty detection methods will then be provided. We will describe the construction and implementation of the three extreme value algorithms first introduced by Vignotto and Engelke (2018) and Rudd et al. (2017). Those three algorithms will be tested on two data-sets along with an additional one-class SVM. Finally, conclusions will be drawn about their respective performances.