Multiview Hybrid Embedding: A Divide-and-Conquer Approach.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Cross-view classification is crucial for matching data from different modalities.
• Multiview subspace learning (MvSL) methods learn shared latent spaces but struggle with nonlinear data and outliers.
• Existing MvSL approaches often fail when data lies on nonlinear manifolds or contains significant noise.

Purpose of the Study:

• To develop a robust and accurate cross-view classification algorithm.
• To address the limitations of existing MvSL methods in handling nonlinearities and outliers.
• To propose a novel approach that enhances discriminative power across different views.

Main Methods:

• A Divide-and-Conquer strategy is employed, breaking cross-view classification into three subproblems.
• Multiview Hybrid Embedding (MvHE) is proposed, with specialized models for view discrepancy removal, nonlinear structure discovery, and discriminability enhancement.
• Kernel extension is applied to further improve the representation capabilities of MvHE.

Main Results:

• MvHE demonstrates superior performance over state-of-the-art MvSL methods on four benchmark datasets.
• The proposed method shows significant improvements in classification accuracy.
• MvHE exhibits enhanced robustness against nonlinear manifolds and data outliers.

Conclusions:

• Multiview Hybrid Embedding (MvHE) offers a powerful solution for cross-view classification, outperforming existing MvSL techniques.
• The method's ability to handle nonlinearities and outliers makes it highly effective for real-world applications.
• MvHE represents a significant advancement in multiview learning for improved classification performance and reliability.