2.9
Anomaly detection
Summary
- This chapter covers core dimensionality-reduction methods for compressing high-dimensional data while preserving useful structure.
- You will compare linear methods (PCA, SVD, LDA) with nonlinear methods (t-SNE, Isomap, Kernel PCA).
- After this chapter, you should be able to choose methods for visualization, denoising, and downstream modeling.
Intuition #
Dimensionality reduction is about preserving the right geometry for your objective: global variance for compression, class separation for supervised projection, or local neighborhoods for visualization.
Detailed Explanation #
What This Chapter Covers #
- statistical and threshold-based anomaly detection for time series
- practical detector behavior under seasonality, trend, and sudden jumps
- implementation patterns for detection, visualization, and validation
What You Can Do After This Chapter #
- choose an anomaly detector based on data behavior and operational constraints
- interpret false positives/false negatives with metric-based diagnostics
- design a reproducible anomaly detection workflow from preprocessing to evaluation