Kurtosis-Guided Denoising Score Matching for Anomaly Detection

A new method called kurtosis-based noise scaling (K-DSM) improves denoising score matching (DSM) for tabular anomaly detection. DSM learns data distributions by training neural networks to recover score functions from noise-corrupted samples, with score magnitude indicating anomaly. The challenge is selecting perturbation scale: too little noise causes unstable estimates, too much erases structure. K-DSM sets per-feature noise levels based on marginal distribution kurtosis, enhancing low-density region coverage and high-density precision without extra complexity. The paper is on arXiv.