Multi-Class Supervised Novelty Detection.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Supervised Novelty Detection (SND) addresses identifying novel patterns in high-dimensional data using labeled information.
• One-Class Support Vector Machines (SVM) are common but struggle with mixed distributions.
• Existing methods face challenges in modeling complex data structures and computational scalability.

Purpose of the Study:

• To develop advanced SVM-like algorithms for improved Supervised Novelty Detection (SND).
• To address the limitations of traditional methods in handling mixtures of distributions.
• To enhance both classification and novelty detection capabilities in high-dimensional settings.

Main Methods:

• Introduced a new class of SVM-like algorithms with a coupling term and l2-norm penalty.
• Formulated the optimization objective in primal and derived a dual Quadratic Programming (QP) formulation.
• Developed a Least-Squares formulation for reduced computational cost and a Pegasos-based formulation for large datasets.

Main Results:

• The proposed methods effectively model mixtures of distributions, outperforming traditional approaches.
• The Least-Squares and Pegasos-based formulations offer significant computational advantages.
• Experimental validation demonstrates the practical importance and usefulness in classification and novelty detection.

Conclusions:

• The novel SVM-like algorithms provide a robust framework for Supervised Novelty Detection (SND).
• The developed formulations offer efficient solutions for high-dimensional data and large-scale applications.
• These advancements contribute to more effective pattern recognition and anomaly detection techniques.