Label Efficient Regularization and Propagation for Graph Node Classification.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Computer Vision

Background:

• Graph convolutional networks (GCNs) are state-of-the-art for semi-supervised node classification.
• Existing methods like GraphHop offer intuitive explanations but lack rigorous mathematical treatment.
• GraphHop's performance is limited by its handling of pseudo-labeled nodes and label aggregation.

Purpose of the Study:

• To propose a Label Efficient Regularization and Propagation (LERP) framework for graph node classification.
• To address the limitations of GraphHop in handling pseudo-labeled nodes and label aggregation.
• To develop an efficient and theoretically guaranteed method for semi-supervised node classification.

Main Methods:

• Developed a novel Label Efficient Regularization and Propagation (LERP) framework.
• Introduced an alternating optimization procedure for solving the LERP framework.
• Proposed the LERP method, which adaptively determines reliable pseudo-labels and refines label aggregation.

Main Results:

• The LERP method consistently outperforms GraphHop and other benchmarking methods across various datasets.
• LERP demonstrates superior performance even at extremely low label rates (1-20 samples per class).
• Theoretical convergence of the LERP method is guaranteed.

Conclusions:

• The LERP framework provides a mathematically rigorous approach to semi-supervised node classification.
• LERP effectively addresses GraphHop's shortcomings, offering improved accuracy and efficiency.
• LERP is a highly effective and computationally efficient method for graph node classification, particularly in low-label regimes.