Noise-guided Unsupervised Outlier Detection
Supplementary Material: zip
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
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.
Keywords: Outlier Detection, Synthetic Outliers, Neural Networks, Unsupervised Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Noise-guided Unsupervised Outlier Detection
Abstract: Over the past decade, we have witnessed enormous research on unsupervised outlier detection techniques, ranging from statistical models to recent deep learning-based approaches. Existing approaches generally limit their discussions to unlabelled data mixed with normal data (inliers) and abnormal data (outliers), constituting only a tiny fraction of the whole value space. Such approaches tend to fall into the local optimum of a specific subspace and hardly generalize to diverse datasets. This paper proposes a novel end-to-end Noise-guided unsupervised Outlier Detector (NOD), which infers the anomaly score of the entire value space via a simple MLP to learn the difference between samples and uniform noise. We further theoretically prove that the learned classifier can separate outliers from inliers under a loose condition. Extensive experiments show that NOD significantly advances UOD performance in 22 diverse real-world datasets by an average of 30.6% ROC_AUC against 11 state-of-the-art counterparts without dataset-specific tuning. The merit is of paramount importance for UOD due to the lack of labeled data for supervision.
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: 4700
Loading