Weight Analysis and Optimization in Fuzzy Modeling

In this paper we propose the use of a set of weights in fuzzy modelling, the class weights, which are assigned to each class of a classification problem. We automatically generate a fuzzy model, using a three-stage methodology: (i) generation of a crisp model from a decision tree, induced from the data, (ii) transformation of the crisp model into a fuzzy one, and (iii) optimization of the fuzzy model’s parameters. Based on this methodology, the generated fuzzy model includes the Θ f parameters, which are all the parameters included in the sigmoid functions. In addition, local, global and class weights are included, thus the fuzzy model is optimized with respect to all these parameters ( Θ f , local, global and class weights). The class weight introduction, which is a novel approach, grants to the fuzzy model the ability to identify the individual importance of each class and thus more accurately reflect the underlying properties of the classes under examination, in the domain of application. The above described methodology is applied to five known medical classification problems, obtained from the UCI machine learning repository, and the obtained classification accuracy is high.