Autonomous Online Multistream Generalization Via Fuzzy Joint Discriminant Analysis

En Yu, Jie Lu, Xiaoyu Yang, Guangquan Zhang

Abstract: Various real-world applications generate multiple interconnected and dynamic data streams that suffer from concept drift, which complicates online decision-making. Existing adaptive learning methods for multistream analysis tend to rely heavily on shared target features, which limit both their generalizability and versatility. To address this research gap, we propose MultiStream Generalization (MSG), a new setting that focuses on integrating knowledge from multiple real-time source streams without the need to access target samples, thereby enabling effective generalization to unseen target streams. To achieve MSG, we propose an Online Multistream Generalization (OMuG) method, which utilizes a sliding window-based online ensemble architecture to accumulate knowledge over time. This framework is coupled with a hierarchical fuzzy system to manage asynchronous drift and inter-stream shift autonomously. More specifically, this hierarchical system incorporates a multi-output Takagi-Sugeno-Kang fuzzy system that learns hidden fuzzy representations through a nonlinear fuzzy mapping in the antecedent phase. These multiple fuzzy IF-THEN rules can address the uncertainty during the learning process and relax the IID assumption violated by asynchronous drift. Subsequently, a novel implementation called Fuzzy Joint Discriminant Analysis (FJDA) transforms the relaxed fuzzy features into discriminative representations, enhancing intra-class compactness and inter-class separability. FJDA also incorporates a constraint to mitigate class imbalance, further improving the model's robustness and generalizability. Experimental results demonstrate that OMuG effectively handles concept drift and enhances generalization across multiple streams, marking significant progress in adaptive learning for dynamic environments.

External IDs:doi:10.1109/tfuzz.2026.3656402