Although two-view Takagi-Sugeno-Kang (TSK) fuzzy classifiers have shown promising performance, most existing methods rely on static fusion and lack dynamic boundary modeling. Moreover, complex coupling architectures may obscure rule-level semantics and reduce interpretability. This study proposes a residual-directed two-view deep TSK fuzzy classifier (R-TSKFC), which establishes a hierarchical two-view collaborative learning framework to overcome the limitations of static fusion and enable adaptive modeling of shared decision boundaries. A two-view training strategy is proposed to implicitly enforce semantic consistency between views and suppress view-specific drift for the parameter space. In the feature space, a residual-directed mechanism explicitly conveys boundary-relevant information between views, allowing the proposed model to focus on under-learned regions and enhance discriminative capability. Extensive experiments on seven UCI benchmark datasets and seven real-world medical datasets demonstrate that R-TSKFC achieves competitive classification performance, with consistent improvements over representative baselines. The model also exhibits competitive performance in terms of classification accuracy, generalization ability, and computational efficiency, while preserving interpretability by relying on original input features. Moreover, the residual-directed mechanism provides an auxiliary perspective for understanding the decision process at each network layer, without compromising the transparency of the learned fuzzy rules.

Hierarchical Takagi–Sugeno–Kang (TSK) fuzzy classifiers combine nonlinear modeling with rule-based decision structures for complex classification tasks. However, prediction errors may propagate through successive layers, and the contributions of different layers require careful consideration when combining their outputs. These limitations motivate a learning strategy that incorporates inter-layer information and adjusts layer contributions. This study proposes a hierarchical TSK fuzzy classifier with adaptive error compensation (DAE-TSK-FC). The adaptive error compensation mechanism incorporates prediction differences between successive layers into subsequent layer inputs while retaining supervision from true labels. A soft relaxation-based output ensemble strategy combines supervised base coefficients with output-dependent adjustments to determine the weights assigned to different layers. Experiments compare DAE-TSK-FC with five benchmark classifiers on six UCI datasets and three real-world rice wine fermentation datasets. The proposed method achieves the highest mean testing accuracies on Iris, FJ-BJ-PCI, and FJ-BQ-PCI, reaching 98.38%, 96.25%, and 92.67%, respectively, and exhibits the smallest testing-accuracy standard deviations on Iris and FJ-BJ-PCI. Moreover, the proposed model preserves the interpretable structure of TSK fuzzy rules, allowing its predictions to be explained through the semantic meanings of fuzzy antecedents and the contributions of rule consequents.