Apprentissage par Regroupement d'Attributs dans les Systèmes d'Inférence Floue

Résumé

Fuzzy Rule-Based Classification Systems are very powerful since they provide an easily interpretable model consisting of linguistic if-then rules. The big challenge of these systems is how to deal with high dimensional databases. Indeed, when the number of attributes is high, an exponential increase of the generated rules number is expected. In this context, we focus on an interesting solution to cope with that problem which is the use of ensemble methods. In these methods, the learning problem which involves a big number of attributes is decomposed into sub-problems of lower complexity. Different classifiers are thus constructed with different projections of the features set and the decisions of the different classfiers are combined in order to form the final classification model. Using an ensemble method, we can take advantage of the following benefits. On the one hand, the complexity of the classification system can notably be reduced. On the other hand, the performance and precision of the classification task can be improved since we take in consideration the opinions of different classifiers rather than one single opinion. Our research lies with the scope of the ensemble methods which are used in the context of fuzzy rule-based classification systems. We are interested in particular in the methods which can be used to regroup the set of attributes into sub-groups of dependent ones. First, we analyze the different methods proposed in the literature to detect associations and correlations between the features. Then, we propose new methods of attributes regrouping which are based on the association rules concept and the frequent itemsets mining. The proposed methods are able to detect interesting associations between the attributes ; these associations can be of different types and shapes. These methods can be applied on numerical as well as categorical attributes. In addition, the proposed methods of attributes regrouping can be applied not only on the fuzzy rule based classification systems, but also on different other learning approaches.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 2

Téléchargements : 0