Multiview Variational Sparse Gaussian Processes.

Area of Science:

• Machine Learning
• Statistical Modeling

Background:

• Gaussian process (GP) models are flexible nonparametric tools for various tasks.
• Variational sparse GP (VSGP) scales GP models to large datasets using inducing points.

Purpose of the Study:

• Extend Variational sparse GP (VSGP) to effectively handle multiview data.
• Develop a novel approach for integrating information from multiple data perspectives.

Main Methods:

• Model each data view with a VSGP, incorporating additional inducing points.
• Couple VSGPs by aligning posterior means at shared inducing point locations.
• Introduce a new GP model in the concatenated feature space to learn these shared inducing points.

Main Results:

• The proposed multiview VSGP (MVSGP) model consistently outperforms single-view VSGP.
• MVSGP demonstrates superior performance compared to state-of-the-art kernel-based multiview baselines.
• Achieved enhanced classification task performance on real-world datasets.

Conclusions:

• The MVSGP model offers a powerful framework for multiview learning.
• This approach effectively leverages shared information across multiple data views.
• MVSGP represents a significant advancement for classification tasks involving complex, multiview data.