Prototype-based Domain Description for one-class classification.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• One-class classification is crucial for anomaly detection and outlier identification.
• Nearest neighbor-based methods are widely used but can be computationally intensive.
• Existing methods like NNDD provide a foundation for domain description classification.

Purpose of the Study:

• Introduce the Prototype-based Domain Description (PDD) rule as a novel one-class classifier.
• Develop efficient algorithms for selecting PDD consistent subsets from training data.
• Evaluate the performance of PDD-based methods on large-scale datasets.

Main Methods:

• The PDD classifier utilizes nearest neighbor distances to a reference set (prototypes).
• Introduced the concept of a PDD consistent subset for optimized classification.
• Developed the CPDD algorithm for approximating minimum size PDD consistent subsets.
• Presented Fast CPDD for efficient handling of very large datasets.

Main Results:

• The CPDD algorithm provides a logarithmic approximation factor for minimum PDD consistent subset computation.
• CPDD significantly reduces subset size compared to CNNDD while maintaining classification quality.
• PDD-based methods demonstrate competitiveness against other one-class classification techniques.
• Fast CPDD proves suitable for classifying large datasets.

Conclusions:

• The PDD classifier offers a robust framework for one-class classification and outlier detection.
• The CPDD algorithm and its Fast CPDD variant provide efficient solutions for large-scale data.
• PDD-based approaches represent a significant advancement in nearest neighbor-based classification.