Concetti Chiave
Cross-conformal anomaly detection methods provide reliable uncertainty quantification for anomaly detection systems, enhancing trust and reducing costs related to false discoveries.
Sintesi
Anomaly detection is crucial in various fields like fraud detection, cybersecurity, and healthcare.
Uncertainty quantification is essential for trustworthy machine learning.
Cross-conformal anomaly detection offers a novel framework for anomaly detection with statistical guarantees.
Different methods like JackknifeAD, CVAD, and CV+AD are introduced for uncertainty-quantified anomaly detection.
The study evaluates the performance of these methods on benchmark datasets.
Results show that cross-conformal methods exhibit higher statistical power and stability compared to split-conformal methods.
Statistiche
"The split-conformal method reliably controls the FDR at the nominal level α = 0.2."
"CV+AD and J+AD exhibit smaller FDR and smaller FDR at the 90th percentile beyond the marginal case."
"Cross-conformal detectors tend to outperform split-conformal detectors regarding statistical power."
Citazioni
"Cross-conformal methods offer a natural and effective addition to the field of conformal anomaly detection."
"Results show that cross-conformal methods exhibit higher statistical power and stability compared to split-conformal methods."