KIR-SMOTE: Knowledge Inconsistency Repair SMOTE for Imbalanced Data Classification

Class imbalance is a pervasive challenge in machine learning, where standard classifiers often exhibit a strong bias toward the majority class. Although the Synthetic Minority Over-sampling Technique (SMOTE) is widely used to address this issue by interpolating new minority samples, it heavily relies on spatial geometry and ignores underlying domain knowledge or logical rules. Consequently, traditional SMOTE frequently synthesizes invalid samples that contain logical contradictions or exacerbate class overlap. To overcome these limitations, this paper proposes a novel over-sampling algorithm termed Knowledge Inconsistency Repair SMOTE (KIR-SMOTE). Instead of strictly avoiding imperfect samples during generation, the core strategy of KIR-SMOTE is to initially allow the creation of "slightly inconsistent samples" to preserve data diversity at the decision boundaries. We establish a formal rule-based knowledge base and define a comprehensive violation index to measure the logical consistency of synthetic samples. Samples exceeding a specific inconsistency threshold are discarded as noise, while those with minor violations are retained and funneled into a dynamic consistency repair mechanism. This mechanism formulates a constrained optimization problem to fine-tune non-core features, minimizing feature perturbation while successfully restoring compliance with domain rules. Synthetic results demonstrate that KIR-SMOTE effectively eliminates semantic contradictions and pseudo-features without sacrificing sample diversity. This approach bridges the gap between geometric interpolation and semantic validity, offering enhanced generalization performance and robust compliance for imbalanced data classification in knowledge-intensive domains such as medical diagnosis and financial fraud detection.