Interactive technique to build fuzzy rule-based systems for classification

Abstract

The construction of classifiers from training data is a current research topic. Most of the techniques involve automatic learning classification models. Models based on machine learning techniques are able to build efficient classifiers, but not always create interpretable models. A visual classifier uses a visual display of data to represent instances of the training set and build, through user interaction with the image, a classification model. This approach has the advantage of providing, to the knowledge engineers, greater understanding of the relationships in the data. This chapter presents a technique for the visual construction of classifiers. The technique uses the parallel coordinate representation to build a set of rules. The membership functions are represented in each coordinate in the traditional way. The flexibility of the technique allows selecting the type of membership functions used and the classification system used by the rules. This chapter presents a group of experimental results that support the validity of the classifiers obtained with this approach. It also presents the dimensions of the sets of rules constructed for multiple datasets. This measure is of great interest in the area of fuzzy rule system to evaluate the interpretability. The experiments were configured to use triangular membership functions and rule systems where the class is selected using a single winner rule. © Springer-Verlag Berlin Heidelberg 2011.