Label as Equilibrium: A performance booster for Graph Neural Networks on node classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Mining

Background:

• Graph Neural Networks (GNNs) are pivotal for node classification tasks.
• Label reuse methods enhance GNN performance but suffer from overfitting and high memory consumption.
• Existing label reuse techniques face limitations in generalization and computational efficiency.

Purpose of the Study:

• To introduce Label as Equilibrium (LaE), an advanced label reuse approach for Graph Neural Networks.
• To address overfitting and memory inefficiency issues in current label reuse strategies.
• To improve the accuracy and generalizability of GNN-based node classification.

Main Methods:

• Developed Label as Equilibrium (LaE) with an improved masking strategy and supervision concealment.
• Implemented a technique for an infinite number of iterations optimizable within constant memory.
• Conducted extensive node classification experiments on eight real-world datasets.

Main Results:

• LaE significantly increased the accuracy scores of prevailing GNNs by an average of 2.31%.
• LaE outperformed previous label reuse approaches by an average of 1.60% across datasets.
• The proposed method demonstrated superior performance and efficiency in node classification tasks.

Conclusions:

• Label as Equilibrium (LaE) effectively resolves overfitting and memory consumption issues in GNN label reuse.
• LaE offers a significant improvement in node classification accuracy and generalizability.
• The LaE approach is broadly applicable and can benefit numerous state-of-the-art GNN models.