Noise-Robust Density Estimation for Tabular Data Anomaly Detection

ICLR 2026 Conference Submission17784 Authors

Keywords: Anomaly Detection

TL;DR: A density estmation method for anomaly detection that isolate the influence of inherent noise inside data.

Abstract: Density-based anomaly detection methods often provide accurate and interpretable predictions but their performance can be severely affected by the inherent noise of data. In this paper, we present a noise-robust density estimation (NRDE) method for tabular data anomaly detection. We aim to estimate density of pure data with influence of noise isolated, which is a non-trivial task since data-generating process is completely unknown. NRDE learns a Jacobian-regularized normalizing flow to estimate the sources of data and categorizes sources into two groups, where one group generates pure data and the other generates noise. Then we can estimate the density of pure data and use it to detect anomalies caused by the sources of pure data rather than the changes caused by the sources of noise. Therefore, compared with other density based methods, our NRDE is much more robust to noise. Besides the new algorithm, we also provide theoretical results to support the effectiveness of NRDE. We compare NRDE with $15$ baselines on $47$ benchmark datasets under different settings, including vanilla anomaly detection, anomaly detection with anomaly contamination, anomaly detection on noisy data and transductive outlier detection. The results demonstrate effectiveness and superiority of NRDE.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17784