Multivariate soft repulsive system identification for constructing rule-based classification systems: Application to trauma clinical data

Rule-based classification systems constructed upon linguistic terms in the antecedent and consequent of the rules lack sufficient generalization capabilities. This paper proposes a new multivariate fuzzy system identification algorithm to design binary rule-based classification structures through making use of the repulsive forces between the cluster prototypes of different class labels. This approach is coupled with the potential discrimination power of each dimension in the feature space to increase the generalization potential. To address this issue, first the multivariate variant of a newly proposed soft clustering algorithm along with its mathematical foundations is proposed. Next, the discriminatory power of each individual feature is computed, using the multivariate membership values in the proposed clustering algorithm to achieve the most accurate firing degree in each rule. The main advantage of this method is to handle unbalanced datasets yielding superior true positive measure while keeping the false positive rate low enough to avoid the natural bias toward class labels containing larger number of training samples. To validate the proposed approaches, a series of numerical experiments on publicly available datasets and a real clinical dataset collected by our team were conducted. Simulation results demonstrated achievement of the primary goals of this research.