Fuzzy Set-based Hybrid Neural Networks Driven with the Aid of Polynomial Neuron/Fuzzy Polynomial Neuron Constructed through Feature Selection

<title>Abstract</title> In this study, we propose an advanced category of fuzzy perceptron architecture—fuzzy set-based hybrid neural networks(FS_HNN) based on feature selection. There is no doubt that there is a great amount of information with different degrees of importance in the raw data. Therefore, in the process of designing regression model, one of the most important missions is to avoid the loss of these significant features. However, regarding the structure of conventional fuzzy polynomial neural network(FPNN), the complexity of each node increases as the number of layers increases in each layer because of the participation of polynomial function. Hence overfitting phenomenon and high-complexity neuron nodes are common in the structure of conventional FPNN. Due to the composition and structure of the conventional FPNN nodes, the original data cannot be retained in the neural network simulation process. Based on these points, we involve the use of the F-test into this study and two types of structure (type I: combined with fuzzy polynomial neurons (FPN) and F-test fuzzy polynomial neurons(F-FPN); type II: combined with fuzzy polynomial neurons and F-test polynomial neurons(F-PN)) were adopted in experiments. Construction of neuron by F-test is focused on saving the essential information coming from the original data, and providing more efficient performance results for follow-up experimental analysis. Furthermore, the proposed model can alleviate the problems of conventional FPNN so that they are effectively optimized with the aid of the F-test feature selection. It is the first successful attempt to draw F-test feature selection technique to fuzzy neural system. The proposed FS_HNN is applied to various datasets and the performance of the proposed model is compared with those obtained by other referred models.