Abstract
Previous studies have shown that the performance of a classifier on imbalanced data heavily relies on informative objects lying in borderline or overlapping areas. In this study, we adopt objective sensitivity to detect the most informative objects based on 0-order Takagi-Sugeno-Kang Fuzzy System (0-TSK-FS), and accordingly propose a novel TSKfuzzy system fusion framework at interpretable sensitivity-ensemble-level (ISE-TSK-FS) to achieve both promising classification performance and reasonable interpretability on imbalanced datasets. Specifically, we assume that a small perturbation of an input object can be interpreted as its neighboring future testing objects. Therefore, object sensitivity w.r.t 0-TSK-FS can be evaluated by the expectation of squared output differences between the input object and the objects in its neighborhood. Being guided by such object sensitivity, we partition the majority class using clustering into different blocks and then construct an ensemble under-sampling process to select the most informative objects from each block iteratively. Moreover, to avoid overfitting issues in assembling, a self-paced factor is introduced to constrain the number of low-sensitive objects but still keep their “skeleton” contribution. Extensive experiments on 7 synthetic datasets, 30 UCI datasets and one real medical case demonstrate ISE-TSK-FS's promising performance and good interpretability on imbalanced data.