Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection

Victor Livernoche, Jie Zan, Reihaneh Rabbany · 2026

Read Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection by Victor Livernoche, Jie Zan, Reihaneh Rabbany (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06955.txt.

Introduction

Victor Livernoche, Jie Zan, and Reihaneh Rabbany improve anomaly detection on table-shaped data using denoising score matching. The method learns a data distribution by training a network to undo added noise, and unusual points stand out afterward. Choosing how much noise to add is the hard part, especially with no labeled anomalies for tuning.

Their scheme sets noise per feature from kurtosis, a measure of a distribution's shape, adapting to sparse and dense regions. A carefully trained single-scale model proves surprisingly strong on standard benchmarks. With a light filtering rule added, it also performs well when the training data itself contain anomalies.

Catalog record

CREATORVictor Livernoche, Jie Zan, Reihaneh Rabbany
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE162 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06955

Browse the full library · About the Tower of Babel