The life cycle of a fuzzy knowledge-based classifier

We describe the life cycle of a fuzzy knowledge-based classifier with special emphasis on one of its most neglected steps: the maintenance of its knowledge base. First, we analyze the process of underwriting Insurance applications, which is the classification problem used to illustrate the life cycle of a classifier. After discussing some design tradeoffs that must be addressed for the on-line and off-line use of a classifier, we describe the design and implementation of a fuzzy rule-based (FRB) and a fuzzy case-based (FCB) classifier. We establish a standard reference dataset (SRD), consisting of 3,000 insurance applications with their corresponding decisions. The SRD exemplifies the results achieved by an ideal, optimal classifier, and represents the target for our design. We apply evolutionary algorithms to perform an off-line optimization of the design parameters of each classifier, modifying their behavior to approximate this target. The SRD is also used as a reference for testing and performing a five-fold cross-validation of the classifiers. Finally, we focus on the monitoring and maintenance of the FRB classifier. We describe a fusion architecture that supports an off-line quality assurance process of the on-line FRB classifier. The fusion module takes the outputs of multiple classifiers, determines their degree of consensus, and compares their overall agreement with that of the FRB classifier. From this analysis, we can identify the most suitable cases to update the SRD, to audit, or to be reviewed by senior underwriters.