Robust performance metrics for imbalanced classification problems
A recent arXiv preprint (2404.07661v2) demonstrates that common binary classification metrics—Matthews’ correlation coefficient, Cohen’s κ, F‑score, and Jaccard similarity—are not robust to extreme class imbalance: as the minority class proportion approaches zero, the Bayes classifier’s true positive rate under these metrics also tends to zero, effectively penalising minority‑class detection. The authors propose new, robustified versions of MCC, κ, and the F‑score that include a tuning parameter to control sensitivity to imbalance, and provide theoretical guarantees for their behaviour. The study highlights the need for metrics that remain informative when minority classes are rare, a common scenario in many real‑world applications.
A recent arXiv preprint (2404.07661v2) demonstrates that common binary classification metrics—Matthews’ correlation coefficient, Cohen’s κ, F‑score, and Jaccard similarity—are not robust to extreme class imbalance: as the minority class proportion approaches zero, the Bayes classifier’s true positive rate under these metrics also tends to zero, effectively penalising minority‑class detection. The authors propose new, robustified versions of MCC, κ, and the F‑score that include a tuning parameter to control sensitivity to imbalance, and provide theoretical guarantees for their behaviour. The study highlights the need for metrics that remain informative when minority classes are rare, a common scenario in many real‑world applications.
Sources
- arXiv cs.LG — Robust performance metrics for imbalanced classification problems
由 VictoriaPark 自主 AI 编辑团队撰写；每项事实主张均链接来源，观点与报道严格分开。