Disentangled Active Learning on Graphs.

Area of Science:

• Graph Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Active learning on graphs (ALG) tackles label scarcity by sampling diverse nodes.
• Current ALG methods often overlook complex latent factors in graph data, limiting sampling effectiveness.
• This can lead to suboptimal node selection and missed opportunities for valuable data acquisition.

Purpose of the Study:

• To introduce Disentangled Active Learning on Graphs (DALG), a novel approach for more effective node sampling in graph learning.
• To address the limitations of existing ALG methods in handling entangled latent factors.
• To improve the efficiency of labeling budgets in graph-based machine learning tasks.

Main Methods:

• Designed the Disenconv-AL layer for learning disentangled feature embeddings.
• Constructed an influence graph for each node and utilized a "memory list" for influence weights.
• Prioritized sampling based on the most significant latent factors from previous rounds to ensure broader coverage.

Main Results:

• DALG achieves superior performance compared to state-of-the-art graph active learning methods.
• Demonstrated improvements of up to approximately 15% in Micro-F1 and Macro-F1 scores.
• Extensive experiments on eight public datasets validate the effectiveness of the proposed approach.

Conclusions:

• DALG offers a new paradigm for active learning on graphs by achieving finer-grained diversity through latent factor disentanglement.
• The method enhances the utility of limited labeling budgets in graph learning scenarios.
• DALG represents a significant advancement in robust and efficient graph-based machine learning.