A unified deep semi-supervised graph learning scheme based on nodes re-weighting and manifold regularization.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Semi-supervised learning on graphs leverages labeled and unlabeled data for predictive modeling.
• Graph Convolutional Networks (GCNs) are effective but have limitations in capturing graph structure and handling topological imbalance.
• Existing GCN methods often overlook manifold structure and graph construction in semi-supervised settings.

Purpose of the Study:

• To propose a novel semi-supervised learning method that overcomes limitations of traditional GCNs.
• To integrate graph learning and graph convolution into a unified network architecture.
• To address topological imbalance and incorporate manifold regularization in graph learning.

Main Methods:

• Introduced ReNode-GLCNMR, a unified network integrating graph learning and graph convolution.
• Enforced label smoothing via an unsupervised loss term.
• Addressed topological imbalance by adaptively reweighting labeled nodes based on class boundary proximity.

Main Results:

• ReNode-GLCNMR significantly outperforms state-of-the-art semi-supervised GNN methods.
• Experimental validation on 8 benchmark datasets demonstrated superior performance.
• The method effectively integrates manifold regularization and adaptive node reweighting.

Conclusions:

• ReNode-GLCNMR offers a robust solution for semi-supervised graph learning.
• The proposed approach enhances predictive accuracy by addressing key GCN limitations.
• This work advances the field of graph-based machine learning and predictive modeling.