Anomaly Detection by Estimating Gradients of the Tabular Data Distribution

Manuel Hirth, Enkelejda Kasneci

Keywords: Anomaly detection, Tabular data, Noise Conditional Score-based Networks

TL;DR: An one-class anomaly detection method by estimating the log-likelihood of the data distribution

Abstract:

Detecting anomalies in tabular data from various domains has become increasingly important in deep learning research. Simultaneously, the development of generative models has advanced, offering powerful mechanisms for detecting anomalies by modeling normal data. In this paper, we propose a novel method for anomaly detection in a one-class classification setting using a noise conditional score network (NCSN). NCSNs, which can learn the gradients of log probability density functions over many noise-perturbed data distributions, are known for their diverse sampling even in low-density regions of the training data. This effect can also be utilized, and thus, the NCSN can be used directly as an anomaly indicator with an anomaly score derived from a simplified loss function. This effect will be analyzed in detail. Our method is trained on normal behavior data, enabling it to differentiate between normal and anomalous behaviors in test scenarios. To evaluate our approach extensively, we created the world's largest benchmark for anomaly detection in tabular data with 49 baseline methods consisting of the ADBench benchmark and several more datasets from the literature. Overall, our approach shows state-of-the-art performance across the benchmark.

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

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11417