Comparison of data clustering methods for automatic determination of granulation in a genetic fuzzy system

The paper proposes applying clustering methods to determine the most appropriate number of fuzzy terms to design a genetic fuzzy system. The fuzzy logic system in this study is used to solve data classification problems and is automatically generated by a genetic algorithm. In this study we used a genetic algorithm with encoding of terms and classes into a binary string, while each individual encoded a rule base. It is necessary to set the number of fuzzy terms parameter to build a rule base, since it significantly affects the quality of the generated classifiers. To determine the best method of data clustering, the most well-known algorithms were compared: DBSCAN, k-means and the mean shift algorithm. To evaluate the efficiency of the selected number of fuzzy terms, computational experiments were performed on several data sets. Based on the results, it was determined that the mean shift algorithm selects such a number of terms that allows building more accurate classifiers in comparison to the other two methods involved in testing. A comparison was also made with alternative classification methods such as k nearest neighbors, support vector machine and neural networks, as a result of which the proposed method showed comparable classification quality. The developed approach to automating the determination of the number of terms makes it possible to exclude manual selection of granulation for various data, reducing the cost of creating an effective fuzzy system for solving classification problems.