Active Learning for Node Classification: An Evaluation.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Deep neural networks require substantial labeled data for training.
• Active learning addresses data scarcity by selecting informative instances for labeling.
• Graph neural networks (GNNs) excel at attributed graph classification but need hyperparameter tuning.

Purpose of the Study:

• To investigate the effectiveness of active learning strategies for attributed graph classification.
• To compare existing active learning algorithms on graph data against their performance on other data types.
• To propose an improved active learning approach for GNNs in label-scarce settings.

Main Methods:

• A comparative analysis of state-of-the-art active learning algorithms on attributed graphs.
• Evaluation of algorithms designed for image and text data when applied to graph structures.
• Development and testing of a novel count-based exploration term within an active learning framework for GNNs.

Main Results:

• Active learning algorithms optimized for images and text demonstrate suboptimal performance on attributed graphs.
• The proposed active learning method, incorporating a count-based exploration term, shows improved efficacy.
• Empirical results on benchmark graphs confirm the necessity of exploration-based strategies for GNN active learning.

Conclusions:

• Standard active learning techniques are not directly transferable to graph neural network applications.
• Integrating exploration-vs.-exploitation considerations is crucial for effective active learning on graph data.
• The novel exploration term enhances uncertainty-based active learning for GNNs, particularly in low-data regimes.