Scalable Semi-Supervised Learning With Discriminative Label Propagation and Correction.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semi-supervised learning utilizes both labeled and unlabeled data but faces challenges with existing methods.
• Current approaches often focus on either similarity structures or regression losses, neglecting their interaction.
• Unreliable similarity structures among boundary samples can mislead label propagation and harm out-of-sample performance.

Purpose of the Study:

• To propose a scalable semi-supervised learning framework, Discriminative Label Propagation and Correction (DLPC), addressing limitations of existing methods.
• To enhance the effectiveness of regression losses for boundary samples by projecting them onto independent class labels.
• To improve label quality and facilitate feature projection learning through collaborative exploitation of regression losses and similarity structures.

Main Methods:

• DLPC collaboratively exploits regression losses and similarity structures.
• Samples are projected onto independent class labels with nonnegative adjustment vectors, enlarging inter-class distances.
• Label propagation occurs through dynamically optimized graph structures, followed by correction using regression losses.

Main Results:

• DLPC effectively improves label quality and facilitates feature projection learning.
• An accelerated solution enhances computational efficiency, making DLPC scalable to large-scale problems.
• Experiments demonstrate DLPC's effectiveness and superiority over state-of-the-art competitors in both single-view and multi-view tasks.

Conclusions:

• DLPC offers a robust and scalable solution for semi-supervised learning by integrating regression losses and similarity structures.
• The framework's ability to handle boundary samples and its computational efficiency make it suitable for diverse applications.
• DLPC shows significant potential for advancing semi-supervised learning research and practical implementations.