Self-Paced Co-Training of Graph Neural Networks for Semi-Supervised Node Classification.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Graph Neural Networks (GNNs) excel in graph data tasks but require extensive labeled data for training.
• Acquiring sufficient labeled data is often impractical and resource-intensive.
• Existing co-training methods in semi-supervised learning (SSL) can be hindered by inaccurate pseudolabels from immature models.

Purpose of the Study:

• To propose a novel Self-Paced Co-Training for GNN (SPC-GNN) framework for semi-supervised node classification.
• To address the issue of inaccurate pseudolabel propagation in traditional co-training approaches.
• To enhance the robustness and performance of GNNs in low-data regimes.

Main Methods:

• The SPC-GNN framework trains multiple GNNs on different data representations, utilizing both labeled and pseudolabeled data.
• A self-paced label augmentation strategy is employed to prioritize high-confidence pseudolabels during training.
• A pretraining step followed by a two-stage optimization scheme is used to train the GNNs effectively.

Main Results:

• The proposed SPC-GNN framework demonstrated significant improvements in node classification tasks.
• The self-paced strategy effectively mitigated the negative impact of inaccurate pseudolabels.
• Experimental results showed superior performance compared to state-of-the-art SSL methods.

Conclusions:

• The SPC-GNN framework offers an effective solution for semi-supervised node classification with limited labeled data.
• Prioritizing pseudolabel quality through self-pacing enhances GNN training stability and accuracy.
• The proposed method represents a significant advancement in SSL for graph-based applications.