A New Approach for Time Series Prediction: Fuzzy Regression Network Functions

Authors : Mehmet Raci Aktoprak, Ozge Cagcag Yolcu

Pages : 36-52

Doi:10.7240/jeps.1573839

View : 84 | Download : 76

Publication Date : 2025-03-25

Article Type : Research Paper

Abstract :The fuzzy regression functions (FRFs) constructs a comprehensive model by combining a series of linear functions based on the inputs. However, the relationship between input and output is not always purely linear. The approach presents novel FRFs with nonlinear structures based on neural networks, combining the strengths of both computational models and fuzzy logic. The proposed model generates membership values by fuzzifying real-valued time series observations, utilizing the fuzzy C-means clustering algorithm. Inputs are then created from the real-valued lagged observations and transformed membership values. A set of feed-forward neural networks, corresponding to the number of fuzzy sets, produces outputs as nonlinear functions of the inputs. These outputs are combined based on the membership values, representing the degree to which each time point belongs to the respective fuzzy sets, to generate the final predictions. The proposed prediction model is referred to as Fuzzy Regression Network Functions (FRNFs). The prediction performance of FRNFs is investigated across several criteria by implementing it in various real-world time series datasets.
Keywords : Bulanık regresyon ağı fonksiyonları, Yapay sinir ağları, Tip-1 bulanık fonksiyonlar, Zaman serisi tahmini