Impact of Data Distribution and Bootstrap Setting on Anomaly Detection Using Isolation Forest in Process Quality...

Area of Science:

• Statistical Process Control
• Machine Learning
• Data Mining

Background:

• Isolation Forest (iForest) is a popular anomaly detection algorithm.
• Its performance with non-normal data distributions and bootstrap resampling is not well-understood.
• Statistical Process Control (SPC) benefits from robust anomaly detection methods.

Purpose of the Study:

• To evaluate the impact of data distribution on iForest anomaly detection performance.
• To assess the effect of bootstrap resampling on iForest.
• To compare iForest against Hotelling's T2 control chart in SPC.

Main Methods:

• A simulation study across 18 scenarios with log-normal, gamma, and t-distributions.
• Varied mean shift levels and bootstrap configurations were tested.
• Performance metrics included accuracy, precision, recall, F1-score, and Average Run Length (ARL)1.

Main Results:

• iForest significantly outperformed Hotelling's T2, particularly with non-Gaussian data and small-to-medium shifts.
• Bootstrap resampling yielded marginal improvements in classification metrics.
• iForest showed reduced sensitivity to subtle process changes (e.g., 1σ mean shift).

Conclusions:

• iForest is a powerful tool for anomaly detection in SPC, especially for non-normal data.
• Bootstrap resampling offers limited benefits.
• Improving iForest's sensitivity to minor process variations is a key area for future research.