Finding rare and atypical events: framing the task without labels, statistical thresholds and robust estimators, Isolation Forest and autoencoders, time series anomalies, fraud detection and production alerting
Foundations & Problem Framing
Anomaly types, supervised vs unsupervised framing, metrics under extreme imbalance, threshold selection, labels from incidents, honest validation and interpretability
Detection Algorithms
Z-score and MAD, extreme value theory, Isolation Forest, LOF, One-Class SVM, autoencoders and VAE, Mahalanobis distance, detector ensembles
Time Series
Seasonality and autocorrelation, STL decomposition, forecast-based detectors, CUSUM and EWMA, changepoint detection, matrix profile, multivariate series, foundation models
Production & Case Studies
Product metric detection service, alerting without noise, threshold calibration, real-time fraud detection, label delay, detector drift, measuring quality in production