Self-Supervised Random Forest on Transformed Distribution for Anomaly Detection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Anomaly detection is crucial for identifying abnormal data points.
• Self-supervised methods using random affine transformations (RATs) show promise but face data overlap issues.
• Effective data distribution is key for successful anomaly detection.

Purpose of the Study:

• Introduce a novel self-supervised forest (sForest) model for enhanced anomaly detection.
• Address the data distribution bottleneck in existing self-supervised methods.
• Improve the stability and effectiveness of anomaly detection algorithms.

Main Methods:

• Leverage random Fourier transform (RFT) to map data into a new feature space.
• Utilize random orthogonal rotations to create a self-labeled training dataset.
• Theoretically prove the stability of the proposed data distribution over RATs.
• Employ a random forest (RF) classifier for anomaly identification.

Main Results:

• The sForest model generates a more stable data distribution compared to RATs.
• Comprehensive experiments demonstrate sForest's superior performance on diverse datasets.
• sForest outperforms various benchmark anomaly detection techniques.

Conclusions:

• The sForest model offers a robust and effective solution for anomaly detection.
• Controlled data distribution via RFT and orthogonal rotations is critical for performance.
• This approach advances self-supervised learning in anomaly detection.