Multi-view learning with enhanced multi-weight vector projection support vector machine.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Multi-view learning utilizes data from distinct feature sets.
• Existing multi-view support vector machine (SVM) methods can be slow and generalize poorly on complex data.
• There is a need for more efficient and effective multi-view learning algorithms.

Purpose of the Study:

• To propose two novel multi-view enhanced multi-weight vector projection support vector machine (MvEMV) models.
• To address the limitations of existing methods regarding processing time and generalization.
• To enhance classification performance in multi-view learning scenarios.

Main Methods:

• Introduced two MvEMV models: ratio form (R-MvEMV) and difference form (D-MvEMV).
• Models generate projection matrices with projection vectors for each view, unlike traditional hyperplane searching.
• Incorporated a co-regularization term to maximize inter-view consistency, simplifying to eigenvalue problems.

Main Results:

• The proposed R-MvEMV and D-MvEMV models demonstrated superior classification performance compared to state-of-the-art methods.
• Numerical tests indicated significantly higher efficiency in terms of processing time.
• The optimal weight vector projections were identified as eigenvectors corresponding to the smallest eigenvalues.

Conclusions:

• The novel R-MvEMV and D-MvEMV models offer a more efficient and effective approach to multi-view learning.
• These methods provide better generalization capabilities on complex datasets.
• The proposed models represent a significant advancement in multi-view support vector machine techniques.