Anomaly detection
Find the points that don't belong. Kinds of anomaly, statistical tests with z-scores and the Mahalanobis distance, distance- and density-based scores (k-NN, LOF), the isolation forest, one-class SVMs, reconstruction error, and how to evaluate a detector when anomalies are rare.
55 min
After this lesson you can
- Distinguish point, contextual and collective anomalies, and outlier detection from novelty detection
- Derive why squared Mahalanobis distances of Gaussian data follow a chi-square distribution, and explain masking
- Compute k-NN distance and local outlier factor scores, and say when local beats global
- Implement an isolation forest from scratch and explain its anomaly score
- Evaluate a detector with rare positives using precision-at-k and average precision, not accuracy
In this lesson
- What counts as an anomaly
- Statistical scores: how many standard deviations?
- Distances and local density
- Isolation: anomalies are few and different
- Four detectors, one dataset
- Evaluating a detector when anomalies are rare
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.