Anomaly-Detection on Ryan Orban

Deep Anomaly Detection with Self-Supervised Learning and Adversarial Training

me@ryanorban.com (Ryan Orban) — Tue, 23 Aug 2022 00:00:00 +0000

Summary

This paper addresses a fundamental problem in anomaly detection: labeled examples of anomalies are scarce or nonexistent in real-world settings, making purely supervised approaches impractical. The work combines self-supervised learning — which extracts useful representations from unlabeled normal data — with adversarial training to build deep anomaly detectors that are both more sensitive to out-of-distribution inputs and more robust to perturbations that might mask anomalies.

Introducing Practical and Robust Anomaly Detection in a Time Series

me@ryanorban.com (Ryan Orban) — Tue, 06 Jan 2015 00:00:00 +0000

Summary

Twitter released AnomalyDetection, an open-source R library for robust time-series anomaly detection, in early 2015. The technical approach combines STL decomposition (Seasonal-Trend decomposition using Loess) with the Generalized ESD test (Extreme Studentized Deviate) — a statistically principled method that handles seasonality, noise, and multiple anomalies in a single pass.

Introduction to One-Class Support Vector Machines

me@ryanorban.com (Ryan Orban) — Sat, 13 Jul 2013 00:00:00 +0000

Summary

Standard Support Vector Machine (SVM) classification draws a decision boundary between two labeled classes. One-class SVMs solve a different problem: you only have examples of one class (normal data), and you want to detect points that fall outside it. This is anomaly detection — or in Schölkopf’s terminology, novelty detection.