Self-configuring hybrid evolutionary algorithm for multi-class unbalanced datasets

This paper describes a modification of the self-configuring hybrid evolutionary algorithm for solving classification problems. The algorithm implements a hybridization of Pittsburg and Michigan approaches, where Michigan part is used together with mutation operator. The rule bases use fixed fuzzy terms, and the number of rules in the rule base can change during the algorithm run. Also, the applied algorithm uses a set of heuristics to determine the weights and class labels for every fuzzy rule, using the confidence values, which are calculated using the training sample. A special initialization procedure allows getting more accurate fuzzy rule bases on the first generations. The modification changes the procedure of determining the most appropriate class number for the fuzzy rule. It uses the number of instances of different classes, as a weighting coefficient to avoid confidence values bias. Also, we apply two classification quality measures, the classical accuracy value and the average accuracy among classes. The modification, combined with different classification quality measures, allows improvement in the classification results. The self-configuring algorithm is tested on a set of unbalanced classification problems with several classes using cross-validation and a stratified sampling procedure. The test problems included image segment classification, bank client classification, phoneme recognition, classification of page contents, and satellite image classification. For one of the problems, the confusion matrixes are provided to show the increasing balance over the class accuracies. The presented method has efficiently solved the satellite images classification problem and can be applied for many real-life problems, including the problems from aerospace area.