Information-controlled graph convolutional network for multi-view semi-supervised classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Graph convolutional networks (GCNs) excel in multi-view learning but suffer from over-smoothing, hindering long-range dependency capture.
• Existing methods to mitigate over-smoothing often sacrifice feature transformation, limiting model expressiveness.

Purpose of the Study:

• Propose an information-controlled GCN for multi-view semi-supervised classification.
• Address the over-smoothing problem while preserving feature transformation capabilities.
• Enhance the stability of GCNs during forward and backward propagation.

Main Methods:

• Impose orthogonality constraints on the feature transformation module to maintain node embeddings during propagation.
• Incorporate a damping factor with residual connections to alleviate over-smoothing.
• Theoretically analyze the model's ability to stabilize forward inference and backward propagation.

Main Results:

• The proposed method effectively alleviates the over-smoothing problem in GCNs.
• Feature transformation is successfully retained, enhancing model expressiveness.
• Experimental results on benchmark datasets validate the proposed approach's effectiveness.

Conclusions:

• The information-controlled GCN offers a robust solution for multi-view semi-supervised classification.
• The method overcomes limitations of previous GCN architectures by balancing over-smoothing mitigation and feature transformation.
• The proposed model demonstrates superior performance and stability in GCN-based learning tasks.