Deep Double Incomplete Multi-View Multi-Label Learning With Incomplete Labels and Missing Views.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view multi-label classification faces challenges with missing views and labels.
• Existing methods often address only one type of incompleteness, not both simultaneously.

Purpose of the Study:

• To propose a novel network capable of handling both missing views and missing labels in multi-view multi-label classification.
• To develop a robust method that effectively utilizes available data and label information despite incompleteness.

Main Methods:

• A network integrating view-specific deep feature extraction, weighted representation fusion, classification, and deep decoding.
• Incorporation of view missing information into the fusion module and label missing information into the classification module.
• Training capability in both supervised and semi-supervised settings for flexible deployment.

Main Results:

• The proposed method effectively reduces the negative impact of missing views and labels.
• It enhances the extraction of discriminative features and improves classifier performance.
• Experimental results on five benchmarks show significant performance gains in both supervised and semi-supervised cases.

Conclusions:

• The developed network provides an effective solution for incomplete multi-view multi-label learning.
• It demonstrates superior classification performance on tasks with both missing views and labels.
• The method offers flexibility through supervised and semi-supervised training options.