A new fuzzy linear regression algorithm based on the simulation of fuzzy samples and an application on popularity prediction of Covid-19 related videos

Abstract A new approach to regression modeling for crisp input-fuzzy output is introduced. The procedure starts with sample generation of symmetrical triangular fuzzy outputs and applying robust linear regression (RLR) a substantial number of times to crisp data. Then, the centers of the coefficients are determined as the mean of upper and lower values. Similarly, the spreads are assumed as the half-length of the resulting intervals. Concurrently, outliers are labeled during the RLR. The total absolute difference between left and right endpoints as a distance between two fuzzy numbers is considered as an error measure. Finally, at the control phase, the estimated spreads are narrowed via bisection. Successively at the correction phase, spreads are widened with respect to outliers, and the constraints, and whether getting a better sum of errors. Numerical examples and comparison studies are given to clarify the proposed method. Furthermore, given the profound effects of the worldwide pandemic, the topic of popularity prediction in YouTube videos related to Covid-19 is chosen as an application.