Unsupervised Real-Time Stream-Based Novelty Detection Technique
summary
An unsupervised, stream-based novelty detector that separates heavy offline training from lightweight online scoring and emits an expected-value vector into the stream. It combines three model types—empirical distributions, spectral nonparametric forecasting, and nearest‑neighbor distances—stores history in a component called "tm series baby," and runs in a scalable node/queue topology. A toy experiment on two VMs with synthetic load achieved a 62.5% detection rate, detecting load spikes but largely missing load drops.
AI-written summary · about
chapters (4)
related talks
43mDetecting Anomalies in Streaming Data
42mDetecting novel anomalies in Twitter
52mDevil in the Haystack
41mScalable Anomaly Detection (with Zero Machine Learning)
30mRealtime Time Series Anomaly Detection in Production
41mApplying Foundational Models for Time Series Anomaly Detection
14mJump-Starting Multivariate Time Series Anomaly Detection for Online Service Systems
12mModel Monitoring: Detecting and Analyzing Data Issues