Chapter 9 #

Anomaly Detection #

Detect rare, interesting, or problematic observations. Combine robust preprocessing, model choice, and thresholding with domain review.

Approaches #

• Statistical: z‑scores, robust IQR, MAD.
• Density/Isolation: Local Outlier Factor, IsolationForest.
• Reconstruction: autoencoders, PCA reconstruction error.
• Forecast residuals (for time series): spikes in model residuals.

Practical notes #

• Standardize features; treat heavy tails before parametric methods.
• Calibrate thresholds with precision‑recall focus; costs are asymmetric.
• Investigate flagged cases to refine features and reduce false positives.