Anomaly detection: Z-score to Autoencoder, class imbalance key

The explainer is written as a technical companion to a separate beginner-oriented piece, and its scope is deliberately broad: it moves from simple statistical scoring like Z-score to tree-based and density-based methods such as Isolation Forest and LOF, then to boundary-based One-Class SVM and reconstruction-based Autoencoder. For time series, it adds ESD and ARIMA residuals rather than treating the problem as a single static snapshot.

The through-line is that no single family covers every case, so the piece recommends combining statistical and machine learning techniques. It also singles out class imbalance as a practical obstacle, a reminder that in real anomaly detection the rare-event nature of the data shapes both method choice and evaluation.

The resource list leans toward G-test certification materials and a free practice-question site, framing it as a study aid for readers building formal credentials alongside hands-on technique.