Online Nonlinear Bagging Using a Neuro-Neo-Fuzzy Approach for Data Stream Mining

Abstract

This paper investigates bagging ensemble models and methods that enable the online aggregation of results from individual ensemble members. The primary objective is to design a computationally efficient metamodel capable of real-time parameter adaptation. To achieve this, the study focuses on developing the architecture of a nonlinear adaptive bagging metamodel, developing adaptive learning methods, and experimentally validating its performance and effectiveness. The research employs neuro-neo-fuzzy systems, the -transform, the Ruspini partition of unity, and adaptive and robust learning methods. The proposed metamodel architecture comprises fuzzification, aggregation, nonlinear-synapse, and tunable-weight layers, followed by an output summing unit. The learning process is implemented via a two-stage procedure: the first stage optimizes the output weights, while the second stage fine-tunes the nonlinear synapse parameters. To enhance robustness and smoothing, forgetting factors are integrated into the learning algorithms, providing a balance between convergence rate and tracking stability. The metamodel was validated using an electricity consumption dataset from Ukraine covering a four-month period. Compared with the three heterogeneous subsystems of the ensemble, the proposed approach achieved higher accuracy, reducing the forecasting error from 3.63% to 3.31%. Computational experiments confirm the system's ability to mitigate noise effectively and adapt to nonstationary demand shifts in real time. Future research will focus on evaluating the metamodel's robustness under high noise levels and its sensitivity to parameter variations.