Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection

Researchers introduce K-DSM, a kurtosis-based noise scaling method for denoising score matching that improves tabular anomaly detection without additional model complexity. The approach achieves state-of-the-art performance by adaptively setting noise levels per feature based on marginal distribution shape, reducing hyperparameter tuning burden in scenarios where anomalies are unknown.

K-DSM addresses a fundamental challenge in anomaly detection systems: balancing noise levels to capture both sparse and dense regions of data distributions. Traditional denoising score matching faces a critical tradeoff where insufficient noise produces unstable estimates in low-density areas, while excessive noise obscures local structure. The kurtosis-guided approach elegantly solves this by tailoring noise scales to individual feature characteristics rather than applying uniform hyperparameters across the dataset.

This work challenges prevailing assumptions in the field that multi-scale or noise-conditioned training architectures are necessary for effective anomaly detection. By demonstrating that a single-scale model with data-adaptive noise scaling can outperform more complex alternatives, the research suggests the community may have over-engineered solutions. The introduction of an EMA-teacher filtering mechanism further enhances performance in contaminated settings, showing practical utility beyond theoretical improvements.

The implications extend to real-world deployment scenarios where validation data containing known anomalies is unavailable. Organizations working with tabular datasets—common in fraud detection, network monitoring, and industrial systems—gain a more practical tool requiring less domain expertise for tuning. The method's simplicity and effectiveness could accelerate adoption of neural score-based anomaly detection in production environments previously reliant on isolation forests or statistical baselines.

Future research should explore how these principles transfer to high-dimensional or mixed-type data, and whether kurtosis-guided scaling benefits other score-matching applications beyond anomaly detection.

• →K-DSM uses per-feature kurtosis-based noise scaling to optimize anomaly detection without increasing model complexity
• →Single-scale models with data-adaptive noise can outperform complex multi-scale architectures in anomaly detection
• →The method achieves state-of-the-art results in semi-supervised settings and strong performance in unsupervised contaminated scenarios
• →Reduces hyperparameter tuning burden when anomalies are unknown and validation sets unavailable
• →EMA-teacher filtering combined with K-DSM enables robust unsupervised anomaly detection in tabular datasets