UniTrain: A universal iterative semi-supervised training framework for graph representation learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph neural networks (GNNs) and graph transformers (GTs) excel at graph-related tasks but struggle in semi-supervised settings due to limited labeled data.
• Over-fitting and task gaps persist despite robust encoders and pre-training, hindering GNN and GT potential.

Purpose of the Study:

• To introduce the Universal iterative semi-supervised Training (UniTrain) framework to enhance semi-supervised learning for GNNs and GTs.
• To address label scarcity and improve node classification performance in graph-based machine learning.

Main Methods:

• UniTrain constructs a semantic graph using hidden vector representations from a pre-training stage.
• It employs label knowledge propagation with uncertainty filtering to infer and refine labels for unlabeled nodes.
• High-confidence pseudo-labels are incorporated to mitigate noise and compensate for limited original label guidance.

Main Results:

• UniTrain significantly improves node classification performance across seven diverse graph benchmarks (Cora, Citeseer, Pubmed, Actor, Cornell, Texas, Wisconsin).
• The framework demonstrates substantial performance gains when applied to both GNNs and GTs.
• Experimental results validate the effectiveness and generalization capability of the UniTrain method.

Conclusions:

• UniTrain effectively enhances fine-tuning in existing self-supervised learning methods for graph data.
• The framework is compatible with any GNN or GT encoder, showcasing broad applicability.
• UniTrain offers a robust solution for improving semi-supervised learning on graphs, particularly in low-label scenarios.