Intelligent information systems are at present actively incorporated in almost all industry fields. Out of many approaches to construction of such systems, the following two methods are worthy of noting: a method involving the use of expert knowledge, which include, in particular, fuzzy modeling), and a method of supervised machine learning, which estimate this knowledge from the available data marked depending on the objectresponse pairs. Each of these methods has its essential advantages and drawbacks. A combined use of both the approaches for solving the classification case is given. The proposed method for training a fuzzy classifier includes fuzzification of the input features of objects, shaping of logical rule conditions for the fuzzy model rule base, and selection of the most typical rule conclusions for filling the relational matrix. The terms of input features and their membership functions behave as the fuzzy model hyperparameters. As an example, the case of binary classification in a two-dimensional coordinate space with a nonlinear distribution of classes is considered. The sets of terms and membership functions ensuring high-quality classification on the training and test data have been found for each of the coordinates. The fuzzy classifier trained using the proposed method is able to solve a number of binary and multi-class classification cases. It can be applied in a situation involving difficulties in specifying an a priori rule base, and supervised learning is allowed to be used as a solution. Since the approach has logical rules at its heart, transparency of the model is ensured, and explanations to the results yielded by the model can be done.
Read full abstract