Co-Labeling for Multi-View Weakly Labeled Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Collecting labeled training data is costly and time-consuming.
• Weakly labeled data and multi-view learning offer potential solutions.
• Existing methods struggle to integrate these effectively.

Purpose of the Study:

• To develop a unified approach for multi-view weakly labeled learning.
• To create robust classifiers using diverse weakly labeled multi-view data.
• To address challenges in semi-supervised learning, multi-instance learning, and relative outlier detection.

Main Methods:

• Proposed a co-labeling approach for multi-view weakly labeled learning.
• Modeled each view as a weakly labeled problem using pseudo-labels from other views.
• Explored strategies for generating pseudo-labels and exploited group structure for multi-layer multiple kernel learning.

Main Results:

• Achieved state-of-the-art performance on text-based image retrieval (NUS-WIDE dataset).
• Demonstrated effectiveness in news classification and text categorization on real-world datasets.
• Validated the approach across multi-view semi-supervised learning, multi-instance learning, and relative outlier detection.

Conclusions:

• The co-labeling approach offers a robust solution for multi-view weakly labeled learning.
• Effectively leverages diverse weakly labeled multi-view data to improve classifier performance.
• Outperforms traditional methods in various complex real-world applications.