A safe semi-supervised graph convolution network.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Graph Convolutional Networks (GCNs) excel with non-Euclidean data but struggle to safely leverage unlabeled data.
• Existing GCN variants often degrade performance by inadequately handling risky unlabeled information.

Purpose of the Study:

• To introduce a novel Safe GCN (Safe-GCN) framework to enhance semi-supervised learning performance.
• To enable the safe and effective utilization of large volumes of unlabeled data in graph-based learning.

Main Methods:

• An iterative labeling process is employed to identify high-confidence unlabeled data points.
• A GCN and its supervised counterpart (S-GCN) are trained iteratively to assign pseudo-labels.
• High-confidence data and pseudo-labels are incorporated into the training set for a final S-GCN model.

Main Results:

• Safe-GCN effectively integrates unlabeled data, improving learning performance.
• The framework demonstrates robust performance on citation network datasets.
• Experimental results confirm Safe-GCN outperforms existing graph-based semi-supervised learning methods.

Conclusions:

• Safe-GCN offers a reliable method for incorporating unlabeled data in semi-supervised learning.
• The proposed framework enhances the capabilities of GCNs by enabling safe exploration of unlabeled data.
• Safe-GCN represents a significant advancement for graph-based semi-supervised learning applications.