Unsupervised Anomaly Detection with Isolation Forest
summary
Isolation Forest is an ensemble of extremely randomized trees implemented in scikit-learn and the claim is concrete: on KDD Cup 99 and SF variants it reaches roughly 99% recall with about 12% false positives, shows stability after ~20 trees and strong sensitivity to the contamination threshold. A 200‑row synthetic AML dataset with embedded red‑flag strategies yields higher anomaly scores for the crafted cases when records are summarized and max_samples is reduced to ~50%, while LOF performs poorly on the same data.
AI-written summary · about
chapters (24)
- 0:00Anomaly detection in AML
- 2:06Defining anomalies
- 3:53AML anomaly detection
- 6:00Money laundering challenges
- 7:22Isolation Forest method
- 9:33Isolation Forest in scikit-learn
- 10:01KDD Cup 99 results
- 13:53Isolation Forest vs LOF
- 14:57Test set results
- 15:20Estimators impact
- 16:25Max samples role
- 16:48Contamination rate
- 18:15AML not simple outlier
- 18:57Synthetic data for AML
- 21:53Testing on synthetic data
- 23:26Folding data for red flags
- 24:49Isolation Forest findings
- 25:12Risk-based AML approach
- 26:15Conclusion
- 26:35Q&A scalability
- 28:33Q&A false positives
- 29:44Q&A metrics
- 30:53Q&A unlabeled data
- 31:56Q&A high false positives
related talks
24mA Walk Through the Isolation Forest
26mAnomaly Detection System for the CERN Cloud
39mDetecting Website Intrusion and Account Compromise with Machine Learning
30mMachine Learning Techniques for Class Imbalances & Adversaries
01h 01mD1T2 - Applying Machine Learning to User Behavior Anomaly Analysis
39mSniffing Out the Strange — A Look at Anomaly Detection
34mML-Bound Anomaly Detection: Experiences Beyond Commodity Observability
35mAnomaly Detection in Smart Buildings using Federated Learning