Posterior misclassification rate-based adaptive composite neighbor oversampling for imbalanced learning

Class imbalance refers to a substantial disparity in sample sizes across classes, which can hinder minority class learning and degrade classification performance. Existing oversampling methods often rely on fixed-size neighborhoods for linear interpolation, making sample generation sensitive to borderline and noisy samples. Therefore, we propose the posterior misclassification rate-guided adaptive composite neighbor synthetic minority over-sampling technique (RAN-SMOTE). First, an adaptive misclassification rate estimation strategy is designed using heterogeneous base learners to characterize the sample distribution and remove noisy samples. Second, the posterior misclassification rate is incorporated to adaptively adjust the neighborhood size and select reliable minority neighbors, thereby reducing the interference from borderline samples. Finally, a probabilistic neighbor selection strategy is developed based on composite scores, and Beta distribution-guided triangular interpolation is employed to generate diverse and reliable synthetic samples. Experiments on 31 imbalanced datasets demonstrate that RAN-SMOTE improves classification performance compared with competing methods. Additional label-noise experiments on five datasets demonstrate that the proposed method exhibits good robustness at noise rates of 5% and 10%.