Disentangling Consistent and Specific Information for Double Incomplete Multi-View Multi-Label Classification

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view multi-label classification (MvMlC) integrates information from diverse sources for sample labeling.
• Real-world MvMlC faces challenges with missing views/labels and extracting robust, consistent, and view-specific representations.

Purpose of the Study:

• To propose a novel framework, Disentangling Consistent and Specific Information (DCSI), for incomplete multi-view multi-label classification.
• To address data incompleteness and improve the extraction of both cross-view consistent and view-specific information.

Main Methods:

• A dual-channel encoder extracts consistent and specific information.
• A view discriminator decouples these information types.
• Dynamic-confidence-aware fusion for consistent representations and equal treatment for specific representations are employed.

Main Results:

• Experimental validation on five datasets demonstrated the effectiveness of the DCSI framework.
• The proposed method outperformed existing state-of-the-art approaches in handling incomplete multi-view multi-label data.

Conclusions:

• The DCSI framework offers a robust solution for multi-view multi-label classification with missing data.
• Disentangling and appropriately fusing consistent and specific information is key to improving MvMlC performance.