Graph Convolution Networks with manifold regularization for semi-supervised learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Vision

Background:

• Graph Convolutional Networks (GCN) are effective for graph-based semi-supervised learning.
• Enhancing label propagation in GCNs is crucial for improving performance.
• Existing methods may have limitations in balancing supervised and unsupervised learning objectives.

Purpose of the Study:

• To introduce a novel Graph Convolutional Network model with Manifold Regularization (GCNMR).
• To enhance the label propagation capabilities of GCNs.
• To improve semi-supervised learning performance through a refined objective function.

Main Methods:

• Proposing GCNs with Manifold Regularization (GCNMR).
• Designing an objective function with supervised and unsupervised terms.
• The supervised term fits predicted labels to known labels.
• The unsupervised term ensures smoothness of predicted labels across all samples.

Main Results:

• The GCNMR model demonstrates superior performance compared to standard GCN.
• The proposed model outperforms other competing graph-based semi-supervised learning methods.
• Experiments on three public image datasets validate the effectiveness of GCNMR.

Conclusions:

• GCNMR offers a powerful enhancement to graph-based semi-supervised learning.
• The model retains GCN advantages while improving accuracy without increased time complexity.
• GCNMR represents a significant advancement in leveraging graph structures for learning.