An interpretable evolving fuzzy neural network based on self-organized direction-aware data partitioning and fuzzy logic neurons

This paper proposes the definition of the architecture of an evolving fuzzy neural network based on self-organizing direction aware data partitioning through stochastic processes based on the dataset used in the model. The choice between four different types of logical neurons in the second layer and the six different types of neuron activation function that compose the artificial neural network of aggregation are defined in a procedure of generating random pairs of combinations between the two factors. The combination that obtains the maximization of training accuracy is chosen to compose the network and perform the composition of the model structure, whose components can be transferred to readable IF-THEN rules for interpretability purposes. Furthermore, the model is able to adapt its parameters and evolve its structure autonomously with new data samples through self-organized direction-aware data partitioning (SODA). In this context, we also propose a technique to measure the degree of changes of neurons (rules), which could be used for structural active learning purposes (e.g, to request user feedback in the case of significant changes). To compare the proposed approach, binary pattern classification tests were performed, and the results were compared with other models of fuzzy neural networks and neural networks, obtaining satisfactory results with the stochastic definition of elements that compose their architecture comparing the final accuracy of the model when classifying real datasets. The proposed model obtained the best result in 3 of the four synthetic bases evaluated, in addition to the best accuracy results in the classification of patterns in five of the nine evaluated real datasets. It highlights the accuracy in problems in patients who underwent breast cancer surgery (72.44%), diabetes evaluation (67.87%), Australian (72.27%) and German (80.51%) credit ratings evaluation, and finally in the classification of radar signals in the ionosphere (90.46%). It is also noteworthy to obtain fuzzy rules in evolution extracted from a real problem of the identification of respiratory diseases through the collection of saliva with 76.67% of accuracy. The results presented by the model were superior to traditional artificial intelligence models, while it was possible to extract knowledge in form of interpretable rules and to realize how these changed over time.