Semisupervised Learning on Graphs With an Alternating Diffusion Process.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semisupervised learning is crucial for systems lacking prior data knowledge.
• Graph-based methods propagate labels but often treat graph construction and label propagation separately.
• Independent management limits exploiting label-graph correlations.

Purpose of the Study:

• To unify graph construction and label propagation into a single framework.
• To enhance label propagation accuracy and robustness by exploiting label-graph correlations.
• To develop an iterative method for simultaneously learning graph structure and labels.

Main Methods:

• Formulated graph construction as a regularized function estimation problem.
• Proposed an alternating diffusion process for iterative optimization.
• Integrated graph learning and label propagation in a unified framework.

Main Results:

• Achieved a dynamic graph adapted to predicted labels.
• Demonstrated superior performance over state-of-the-art methods.
• Showcased robustness on synthetic and real-world datasets.

Conclusions:

• The unified framework effectively integrates graph construction and label propagation.
• Iterative refinement leads to more accurate and robust semisupervised learning.
• The proposed alternating diffusion process offers a powerful approach for agnostic learning settings.