Multi-View Deep Gaussian Processes for Supervised Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Multi-view learning aims to improve prediction by integrating information from multiple sample perspectives.
• Existing multi-view deep Gaussian processes excel in unsupervised tasks but are limited for supervised learning and uncertainty estimation.
• Jointly exploring diverse information across multiple views remains a significant challenge in machine learning.

Purpose of the Study:

• To propose a supervised multi-view deep Gaussian process (SupMvDGP) model for labeled multi-view data.
• To enhance prediction performance by leveraging view labels and enabling quantitative uncertainty estimation.
• To develop a model capable of establishing asymmetric depth structures to accommodate view diversity.

Main Methods:

• Introduced the supervised multi-view deep Gaussian process (SupMvDGP) model.
• Employed variational inference for efficient model optimization.
• Designed an asymmetric depth structure to model diverse views effectively.

Main Results:

• SupMvDGP achieved state-of-the-art performance across multiple real-world datasets and tasks.
• The model demonstrated superior effectiveness and superiority compared to alternative deep models.
• A case study confirmed SupMvDGP's capability for robust uncertainty estimation.

Conclusions:

• The proposed SupMvDGP model significantly advances supervised multi-view learning.
• SupMvDGP offers valuable quantitative uncertainty estimation to aid decision-making in high-risk applications.
• The model's ability to handle view diversity and improve prediction performance is validated.