Neuro-Fuzzy Random Vector Functional Link Neural Network for Classification and Regression Problems

Abstract

The random vector functional link (RVFL) neural network has shown the potential to overcome traditional artificial neural networks' limitations, such as substantial time consumption and the emergence of suboptimal solutions. However, RVFL struggles to provide comprehensive insights into its decision-making processes. We propose the Neuro-fuzzy RVFL (NF-RVFL) model by combining RVFL with neuro-fuzzy system. The proposed NF-RVFL model takes humanlike decisions based on the IF-THEN approach and enhances its transparency in decision-making. Within this framework, input features undergo a fuzzification process as they traverse the fuzzy layer. The resulting fuzzified features then navigate a hidden layer through random projection as well as yielding defuzzified values via defuzzification. The defuzzified values, hidden layer outputs and original input features collectively contribute to the output prediction process. The proposed NF-RVFL model employs three distinct clustering methods to establish fuzzy layer centers: randomly initialized centers (referred to as R-means), K-means clustering centers, and fuzzy C-means clustering centers. This approach generates three distinct model variations, namely NF-RVFL-R, NF-RVFL-K and NF-RVFL-C, each producing a diverse set of fuzzified and defuzzified samples. Our research involves experiments on various UCI benchmark datasets, covering binary, multiclass classification, and regression tasks. The statistical tests and comprehensive experimental analyses consistently show that all variations of the proposed NF-RVFL model outperform baseline models, highlighting their generalization capabilities. The proposed NF-RVFL models show the generic nature by being adeptly applicable and excelling in regression as well as classification tasks.