C-parameter version of robust bounded one-class support vector classification.

Area of Science:

• Machine Learning
• Computer Science
• Data Mining

Background:

• One-class classification and anomaly detection are crucial for identifying unusual patterns.
• Existing ν-one-class support vector classification (ν-OCSVC) methods may lack a unique decision boundary and are sensitive to data contamination.
• Outliers and mislabeled data can compromise the performance of traditional one-class classifiers.

Purpose of the Study:

• To propose a novel C-parameter bounded one-class support vector classification (C-BOCSVC) method that ensures a unique decision boundary.
• To develop a robust version (C-RBOCSVC) that enhances resistance to noise and outliers in training data.
• To theoretically analyze the proposed methods and validate their effectiveness through empirical evaluation.

Main Methods:

• Introduced C-BOCSVC utilizing an L1-norm regularization for structural risk minimization, defining a geometric margin in an augmented space.
• Developed C-RBOCSVC by incorporating k-nearest neighbor relative density for adaptive observation weighting, mitigating outlier influence.
• Derived theoretical properties including primal-dual relationships, connections to ν-OCSVC, and computational complexity analysis.

Main Results:

• Experimental results on large datasets confirm the feasibility and reliability of the proposed C-BOCSVC method.
• The C-RBOCSVC demonstrated superior performance over state-of-the-art one-class classifiers when dealing with contaminated datasets.
• The proposed methods effectively address the limitations of existing approaches in handling noisy and outlier-ridden data.

Conclusions:

• The novel C-BOCSVC and its robust variant C-RBOCSVC offer significant improvements for one-class classification and anomaly detection tasks.
• C-RBOCSVC provides enhanced anti-noise and anti-outlier capabilities, making it suitable for real-world applications with imperfect data.
• The publicly available demo code facilitates further research and application of these advanced classification techniques.