Emilie Dumont Email and Phone Number

Lab session in Multimedia Information Retrieval : Text Indexing and Searching and Content Based Image RetrievalLab session in Intelligent Systems : Genetic algorithm, Neural Networks and Decision Trees for Intrusion DetectionLab session in Software Development: Méthode d'ordonnancement and Script of Shell-Unix

Boosting methods constitute a family of algorithms of machine learning which build models (of classification or regression) based on the combination of weak learning. The success of this paradigm, has been extensively studied over the last decade, not only comes from the simplicity of implementation of the methods which implement it and excellent performances of generalization which they make it possible to reach but also theoretical framework well defined in which these methods can be justified. Within the framework of classification, AdaBoost is one of the most popular algorithms and, in spite of its simplicity, one of the most powerful classifiers to date. However, these good performances are called into question when AdaBoost is applied to data containing outliers. We proposed an original variant of regularized AdaBoost making it possible to estimate the posterior probability for example to be an outlier.