Hybrid Algorithm for Tuning Feature Weights in a Fuzzy Classifier

摘要

The paper proposes the use of a hybrid optimization algorithm to tune the weighting coefficients of features in order to increase the quality of fuzzy classifiers in processing imbalanced data. Tuning the weights is intended to adjust the importance of the features in the rule base. A method for calculating fuzzy inference taking into account weight coefficients is proposed. The hybrid is based on a combination of two metaheuristics: gravitational search algorithm and shuffle frog leaping algorithm. It operates in a continuous mode and searches for a vector of weights that maximizes the mean geometric accuracy of the classifier. Experimental results showed improvement in geometric mean accuracy for 33 of 36 imbalanced data sets with two classes compared to the basic fuzzy classifier constructed using the algorithm based on extreme values of features in classes. Although the stage of tuning weights is inferior in efficiency to the stage of optimization of terms, there are tools to improve it further.