Isolation Forest Detects Anomalies Without Modeling Normal or Optimizing Anything

Isolation Forest is an anomaly detection algorithm that identifies outliers not by modeling normal behavior, but by measuring how few random cuts it takes to isolate a data point from the rest. Unlike K-Means or DBSCAN, it skips loss functions and optimization entirely, relying on pure randomness averaged across hundreds of trees to produce an anomaly score. Counterintuitively, the algorithm deliberately trains each tree on a small subsample — typically just 256 points — not as a speed compromise but because smaller samples keep rare points proportionally rare, sharpening isolation. The algorithm outputs a ranked score per point rather than a binary decision, with a user-set 'contamination' parameter estimating the expected fraction of anomalies to derive the classification threshold. The core insight is that rarity is something randomness naturally exposes, making cheap random averaging more effective than careful statistical modeling for this specific task.