A robust multi-view support vector machine with the RoBoSS loss function

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view learning leverages multiple data representations for improved performance.
• Existing support vector machine (SVM)-based multi-view models often neglect complementary information and lack robustness to noise and inconsistencies.
• There is a need for robust multi-view learning methods that can effectively integrate diverse data representations.

Purpose of the Study:

• To propose a novel robust multi-view SVM framework, RoBoSS-MvSVM, that addresses limitations of existing methods.
• To explicitly integrate both consensus and complementary information across multiple views.
• To enhance the robustness and generalization performance of multi-view learning.

Main Methods:

• Developed a robust multi-view SVM framework (RoBoSS-MvSVM) utilizing the RoBoSS loss function.
• The RoBoSS loss function is designed for robustness, boundedness, sparsity, and smoothness, handling noisy and inconsistent samples.
• Optimization is performed using the Nesterov accelerated gradient algorithm, with generalization capacity analyzed via Rademacher complexity.

Main Results:

• RoBoSS-MvSVM demonstrated superior performance compared to baseline methods across synthetic, UCI/KEEL, and Animal with Attribute datasets.
• The proposed method effectively integrates consensus and complementary information, leading to enriched data representation and resilient learning.
• Experimental results consistently showed the outperformance of RoBoSS-MvSVM, with stability confirmed by hyperparameter analysis.

Conclusions:

• RoBoSS-MvSVM offers a robust and effective solution for multi-view learning by integrating diverse information and handling data imperfections.
• The framework provides reliable generalization performance, validated through theoretical analysis and comprehensive experiments.
• The proposed method represents a significant advancement in robust multi-view SVM techniques.