From Local to Global: A Progressive Collaborative Learning Framework for Multitask TSK Fuzzy System Modeling

Multitask Takagi–Sugeno–Kang fuzzy systems (MT-TSK-FS) commonly utilize interpretable fuzzy rules to facilitate information sharing among tasks and have shown promising performance. However, existing MT-TSK-FS modeling approaches still face several challenges. Specifically, they mainly focus on the global fitting of tasks while overlooking the local fitting accuracy of individual rules, thereby undermining rule interpretability. Moreover, the antecedent and consequent components of fuzzy rules are often learned independently, lacking collaborative coordination. In addition, task-specific rule diversity is insufficiently addressed, which may result in redundant or overlapping rules. To this end, we propose a progressive collaborative learning framework for MT-TSK-FS (MTTSKFS-PCL), which progressively optimizes the model from the local perspective of individual rules to that of global tasks. At the local level, a locally target-guided multitask fuzzy clustering method is introduced to enhance local rule accuracy and to jointly optimize rule antecedents and consequents. At the global level, a mini-batch gradient descent (MBGD) algorithm is employed to integrate fuzzy rules and enhance overall task fitting. Meanwhile, a confidence-supervised regularization strategy is introduced to preserve local rule accuracy throughout the MBGD process. Furthermore, a rule diversity enhancement mechanism is incorporated to improve task-specific rule diversity. Extensive experimental evaluations on multiple benchmark datasets demonstrate that the proposed MTTSKFS-PCL significantly outperforms existing state-of-the-art methods, validating its effectiveness and robustness in multitask fuzzy modeling.