Ensemble Teaching for Hybrid Label Propagation

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Label propagation algorithms diffuse labels across similarity graphs.
• Current methods lack stability and ignore example labeling difficulty.

Purpose of the Study:

• Introduce Hybrid Diffusion under Ensemble Teaching (HyDEnT) for improved label propagation.
• Enhance stability and accuracy in semi-supervised learning.

Main Methods:

• Integrate multiple propagation methods as 'learners'.
• Employ an ensemble of 'teachers' to guide propagation using a curriculum.
• Solve an optimization problem via block coordinate descent to select examples.

Main Results:

• HyDEnT demonstrates enhanced stability and consistent performance.
• Propagation proceeds logically from simple to difficult examples.
• Achieved highest classification accuracy on six datasets compared to state-of-the-art methods.

Conclusions:

• HyDEnT offers superior label propagation quality over existing techniques.
• The ensemble teaching approach effectively addresses labeling difficulties.
• The proposed method provides a stable and accurate solution for semi-supervised learning.