Online support vector machine based on convex hull vertices selection.

Area of Science:

• Machine Learning
• Computational Intelligence
• Data Science

Background:

• Support Vector Machines (SVM) are efficient classification techniques but struggle with online learning due to retraining requirements.
• Retraining SVMs with new misclassified samples is computationally expensive and time-consuming for real-time applications.

Purpose of the Study:

• To develop an effective online SVM classifier that overcomes the limitations of traditional SVM retraining.
• To propose the VS-OSVM algorithm for efficient online classification by leveraging geometric properties of SVM.

Main Methods:

• The VS-OSVM algorithm employs a two-step process: skeleton sample selection and online updating.
• Skeleton samples, approximating convex hull vertices within each class, are selected to represent the dataset.
• The classifier is updated using newly arrived samples and the pre-selected skeleton samples.

Main Results:

• Theoretically, the first d+1 selected samples are proven to be convex hull vertices, preserving maximal information.
• Experimentally, VS-OSVM demonstrates effective online classification without performance degradation.
• Benchmark dataset results validate the efficiency and effectiveness of the proposed VS-OSVM algorithm.

Conclusions:

• VS-OSVM provides a viable solution for online classification problems previously intractable for standard SVMs.
• The convex hull vertices selection method ensures efficient and accurate online model updates.
• This approach significantly reduces computational overhead while maintaining high classification accuracy in dynamic environments.