Evidential Bagging: Combining Heterogeneous Classifiers in the Belief Functions Framework

Archive ouverte

Sutton-Charani, Nicolas | Imoussaten, Abdelhak | Harispe, Sébastien | Montmain, Jacky

Edité par CCSD -

International audience. In machine learning, Ensemble Learning methodologies are known to improve predictive accuracy and robustness. They consist in the learning of many classifiers that produce outputs which are finally combined according to different techniques. Bagging, or Bootstrap Aggre-gating, is one of the most famous Ensemble methodologies and is usually applied to the same classification base algorithm, i.e. the same type of classifier is learnt multiple times on bootstrapped versions of the initial learning dataset. In this paper, we propose a bagging methodology that involves different types of classifier. Classifiers' probabilist outputs are used to build mass functions which are further combined within the belief functions framework. Three different ways of building mass functions are proposed; preliminary experiments on benchmark datasets showing the relevancy of the approach are presented.

• Description
• Sujet/Public
• Outil
• Outil d'anticipation
• Mémoire et thèse
• Gestion