Data-driven partitioning approaches for type-2 fuzzy set based time series

Abstract

Fuzzy Set Based Time Series (FTS) prediction techniques offer potential advantages in efficient and intuitive data partitioning and the effective handling of uncertainty in the data. However, such prediction models are commonly challenging to design, requiring careful and application-specific tuning of hyperparameters to provide competitive forecasting performance. Conventional FTS models often rely on predefined partitioning schemes and user-specified hyperparameters, which may introduce subjectivity and limit their adaptability to complex datasets. This thesis presents novel data-driven interval type2 fuzzy set based time series models — SODA-T2FTS, ADP-T2FTS, and ADP-T2LIMG — that integrate interval type-2 fuzzy logic with advanced partitioning techniques, namely the Self-Organized Direction Aware Data (SODA) and Autonomous Data Partitioning (ADP) algorithms. These approaches handle epistemic uncertainty and improve predictive performance while reducing reliance on user intervention. The models were evaluated on financial, benchmark, and energy time series datasets, and evaluation was overall performed in terms of the average number of rules (c.f. interpretability), error metrics, execution time, model complexity and noise sensitivity. Results showed that the proposed models presented superior accuracy, lower forecasting errors, and competitive interpretability compared to state-of-the-art forecasting techniques, also highlighting that data-driven approaches significantly enhance fuzzy set based time series forecasting by improving partitioning accuracy, reducing subjectivity, and increasing adaptability to different datasets.