Semisupervised Prior Free Rare Category Detection With Mixed Criteria

Area of Science:

• Data Science
• Machine Learning
• Anomaly Detection

Background:

• Rare category detection addresses finding statistically significant anomalies in skewed datasets.
• Existing methods often require prior knowledge of the total number of classes, limiting real-world applicability.

Purpose of the Study:

• To propose a novel, prior-free framework for rare category detection.
• To enhance anomaly detection in skewed data distributions using active and semisupervised learning.

Main Methods:

• A new framework combining active learning and semisupervised hierarchical density-based clustering.
• The approach is designed to be prior-free, meaning it does not require pre-specified class counts.

Main Results:

• The proposed method demonstrates improved performance in identifying rare classes.
• It effectively handles datasets with nonlinear mappings and sophisticated class boundaries.
• Superior results were observed on both synthetic and real-world datasets compared to existing techniques.

Conclusions:

• The developed framework offers a flexible and effective solution for rare category detection.
• Its ability to work without prior class information makes it suitable for diverse, real-world scenarios.
• The method enhances the detection of anomalies in complex data structures.