Decision-boundary aware undersampling via Adversarial Influence for imbalanced software defect prediction
Abstract
Class imbalance severely impairs the capacity of software defect prediction (SDP) models to distinguish between defective (minority class) and non-defective (majority class) software instances. Conventional undersampling methods alleviate this problem by reducing the number of non-defective instances; however, they frequently eliminate informative non-defective instances essential for constructing a reliable decision boundary, thereby compromising prediction performance. The key challenge lies in determining which non-defective instances should be retained to preserve an effective decision boundary while balancing class distributions. To address this, the authors propose Adversarial Influence-Based Intelligent Undersampling (AIIUS), a decision-boundary-aware undersampling method for imbalanced SDP. AIIUS introduces the Adversarial Influence metric to quantify the importance of non-defective instances to the model's decision boundary and formulates non-defective instance selection as an Adversarial Influence-based ranking problem, computing Adversarial Influence scores via adversarial perturbations on a surrogate model to select the top-ranked instances most critical to the decision boundary. Experiments on 18 public defect datasets and five commonly used classifiers show that AIIUS consistently improves imbalance-robust performance, achieving average gains of 9.50% in Balance, 3.43% in Area Under the ROC Curve (AUC), and 17.02% in Matthews Correlation Coefficient (MCC) over representative resampling baselines, with statistically significant improvements in most settings.