m-SNE: Multiview Stochastic Neighbor Embedding.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Dimension reduction is crucial for tasks like image retrieval and document classification.
• Conventional feature concatenation struggles with heterogeneous multiview data, especially when noisy.
• Each data view possesses unique statistical properties and interpretations, necessitating specialized handling.

Purpose of the Study:

• To propose a novel multiview stochastic neighbor embedding (m-SNE) method.
• To systematically integrate heterogeneous features into a unified representation using a probabilistic framework.
• To automatically learn optimal combination coefficients for each data view.

Main Methods:

• Developed a probabilistic framework for multiview data integration.
• Introduced an algorithm that learns view-specific combination coefficients.
• The coefficient learning algorithm converges optimally at a rate of O(1/k(2)).

Main Results:

• m-SNE effectively integrates heterogeneous features from multiple data views.
• Learned combination coefficients leverage complementary information across views.
• Demonstrated effectiveness and robustness on synthetic and real-world datasets.

Conclusions:

• m-SNE offers a superior approach to dimension reduction for multiview data compared to conventional methods.
• The learned coefficients enhance the utilization of complementary information, improving performance.
• m-SNE is effective for various applications including data visualization, image retrieval, and object categorization.