Linear regression-based efficient SVM learning for large-scale classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Additive kernels achieve high accuracy in large-scale image classification.
• Current challenges include slow learning speeds and scalability for Support Vector Machines (SVMs).

Purpose of the Study:

• To develop efficient algorithms for large-scale SVM classification utilizing additive kernels.
• To address the limitations of existing methods in terms of speed and scalability.

Main Methods:

• Proposed a linear regression SVM framework to approximate gradient computations for nonlinear kernels.
• Introduced the Power Mean SVM (PmSVM) algorithm employing nonsymmetric explanatory variable functions for additive kernels.
• The PmSVM method avoids closed-form Fourier transforms and extra training for kernel approximation.

Main Results:

• PmSVM demonstrated superior learning speed and accuracy compared to recent algorithms on large-scale benchmark datasets.
• Performance was evaluated on datasets with millions of examples and dense feature dimensions.
• The proposed methods effectively handle large-scale classification tasks.

Conclusions:

• The PmSVM algorithm offers a significant advancement for large-scale SVM classification with additive kernels.
• The novel approach enhances both computational efficiency and predictive performance.
• This work contributes to overcoming scalability challenges in machine learning.