Dual Contrastive Prediction for Incomplete Multi-View Representation Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Incomplete multi-view representation learning presents challenges in unifying diverse data sources and handling missing information.
• Existing methods often address view consistency and data recovery separately, limiting overall performance.

Purpose of the Study:

• To propose a unified information-theoretical framework for simultaneous cross-view consistency learning and missing view recovery.
• To develop a novel objective function that achieves a provably sufficient and minimal representation for incomplete multi-view data.

Main Methods:

• Utilizing contrastive learning to maximize mutual information between different views for consistency.
• Employing dual prediction to minimize conditional entropy for recovering missing views.

Main Results:

• The proposed method integrates consistency learning and data recovery into a single, theoretically grounded approach.
• Demonstrated superior performance over 20 existing multi-view learning methods on six datasets for clustering, classification, and action recognition.

Conclusions:

• This work is among the first to theoretically unify cross-view consistency and data recovery in representation learning.
• The novel framework significantly advances the field of incomplete multi-view learning, offering a robust solution for real-world applications.