Improved multi-view GEPSVM via Inter-View Difference Maximization and Intra-view Agreement Minimization.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview Generalized Eigenvalue Proximal Support Vector Machine (MvGEPSVM) is a recent advancement in multiview data classification.
• Existing MvGEPSVM methods often overlook inter-view discrimination and intra-view agreement, lacking robustness guarantees.

Purpose of the Study:

• To propose an improved Multiview Generalized Eigenvalue Proximal Support Vector Machine (IMvGEPSVM) for more effective multiview data classification.
• To enhance discrimination between different views and agreement within the same view.
• To improve the robustness of multiview classification models.

Main Methods:

• Introduced a multi-view regularization to connect different views of the same class.
• Incorporated maximization of samples from different classes in heterogeneous views to promote discrimination.
• Employed L1-norm for distance calculation to mitigate the impact of outliers, enhancing model robustness.
• Developed an efficient iterative algorithm to solve the objective function.

Main Results:

• The proposed IMvGEPSVM method demonstrates improved effectiveness in multiview data classification.
• The multi-view regularization enhances discrimination and agreement across views.
• The use of L1-norm effectively reduces the influence of outliers, leading to a more robust model.
• The presented iterative algorithm is efficient and its convergence is theoretically proven.

Conclusions:

• The IMvGEPSVM method offers a significant improvement over existing techniques for multiview classification.
• The enhanced regularization and outlier handling contribute to more accurate and robust classification performance.
• The theoretical convergence proof validates the reliability of the proposed algorithm.