Contribution to automatic design of a hierarchical Fuzzy Rule Classifier

In this paper, two ways for automatically designing a hierarchical classifier is checked. This study deals with a specific context where is necessary to work with a few number of training samples (and often unbalanced), to manage the subjectivity of the different output classes and to take into account an imprecision degree in the input data. The aim is also to create an interpretable classification system by reducing its dimensionality with the use of Feature Selection and Fuzzy Association Rules generation. The obtained results over an industrial wood datasets prove their efficacy to select input feature and they are used to make some conclusions about their performance. Finally, an original methodology to automatically build a hierarchical classifier is proposed by merging the both previous methods. Each node of the hierarchical structure corresponds to a Fuzzy Rules Classifier with selected inputs and macro classes for output. The leaves are the outputs of the classification system.