Locally linear SVMs based on boundary anchor points encoding.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Support Vector Machines (SVM) are powerful classification tools.
• Linear SVMs are efficient but limited to linearly separable data.
• Kernel SVMs handle non-linear data but are computationally expensive.

Purpose of the Study:

• To propose a novel classifier, locally linear classifier based on boundary anchor points encoding (LLBAP).
• To achieve the efficiency of linear SVM and the classification power of kernel SVM.
• To develop a method that efficiently classifies linearly non-separable data.

Main Methods:

• LLBAP partitions data into approximately linearly separable subsets using boundary point scanning and local coding.
• Each subset is then classified using a standard linear SVM.
• This approach leverages the strengths of both linear and kernel methods.

Main Results:

• Experiments on large-scale benchmark datasets show LLBAP is more efficient than kernel SVM in training and testing.
• LLBAP demonstrates superior efficiency and classification accuracy compared to other locally linear classifiers.
• The method successfully handles linearly non-separable data.

Conclusions:

• LLBAP offers a computationally efficient and accurate alternative for classifying complex datasets.
• The proposed method effectively bridges the gap between linear and kernel SVMs.
• LLBAP shows significant promise for large-scale machine learning applications.