In the thesis, we provide two main contributions addressing the Feature Selection task, based on the conceptual tools of power index derived from Coalitional Game Theory.
Both contributions are based on a real-world use case rely on fraud alarm generation. We observe that typically the aggregation of the rules in their process of the governance of a pool of Near-Real-Time classification rules are operated in OR. Thus, it is remarkably non-additive, and we argue that in a non-additive context, assessing the performance in isolation might not guarantee optimal performance of the ensemble. We propose in a first time to support the rule management process and then we suggest an approach to the assessment of specific rules
based on their contribution to the overall pool performance. Our method ranks the rules based on their Shapley Value with respect to a target pool performance metrics and the top-k rules are kept operational, while the reminders are considered for deletion from the pool. The second contribution is divided into two parts and aimed at self-training technique (the classifier is first trained on labelled data and the result is used to annotate unlabeled data; then unlabeled data classified with high threshold confidence are added to the training set).
This technique uses its predictions to improve at each iteration. In the first part, deal with the well-known power indices Shapley value, Banzhaf index we also propose a new power index, the restricted Banzhaf index/ k-Banzhaf index an improved version of Banzhaf index. This new power index was obtained in collaboration with some members of the team.
In the second part, we use the three power indices studied and two greedy-based methods, apply them to the semi-supervised machine learning technique, with the scope to improve the quality of our feature selection process. 

Keywords: Détection de Fraud à la Carte, Théorie des Jeux de Coalition, Indice de Pouvoir, Valeur de  Shapley, Indice de  Banzhaf, Indice de Banzhaf restreint, Apprentissage Semi-supervisé, Apprentissage supervisé , Auto-apprentissage.