C6Anomaly detection

Anomaly detection is the search for the rare and the wrong: fraud, defects, intrusions, sensor faults, the handful of points that do not look like the rest. The defining difficulty is that anomalies are scarce and often unlabeled, so you usually model what normal looks like and flag whatever falls outside it, rather than training a standard classifier.

You will begin with statistical baselines (IQR, ESD, and CUSUM thresholds) that are simple and easy to defend, then build Isolation Forest, which isolates outliers in few random splits with no distance metric required. From there you will carve out a region of normal data with a one-class SVM and use its boundary as a novelty detector, and finally train an autoencoder on normal examples and flag points whose reconstruction error runs high.

These techniques run in fraud and security pipelines, manufacturing quality control, and monitoring systems, and the reconstruction-error idea connects directly to deep autoencoders and to drift detection in production ML.