LaenNet: Learning robust GCNs by propagating labels.

Area of Science:

• Graph representation learning
• Machine learning
• Artificial intelligence

Background:

• Graph Convolutional Networks (GCNs) are powerful for graph representation learning but struggle with real-world noisy and sparse data.
• Imperfect graph data, including noisy/sparse features or labels, challenges the robustness of existing GCN models.
• There is a need for GCN architectures that can effectively handle data imperfections.

Purpose of the Study:

• To propose a novel architecture, LAbel-ENhanced Networks (LaenNet), to enhance the robustness of GCNs.
• To improve GCN performance on graph data with noisy or sparse features and labels.
• To provide a generalizable module that can be integrated into various GCN variants.

Main Methods:

• Introduced the LaenNet module, designed to propagate labels concurrently with features within GCNs.
• Integrated the LaenNet module into a hidden layer of GCNs to combine propagated labels with hidden representations.
• Evaluated LaenNet on semi-supervised node classification tasks using datasets with four types of data imperfections: noisy features, sparse features, noisy labels, and sparse labels.

Main Results:

• LaenNet demonstrated superior performance and robustness compared to state-of-the-art baseline models across all tested noisy and sparse graph data scenarios.
• The proposed method effectively handles various forms of data imperfections in graph representation learning.
• Empirical results validate the effectiveness of integrating label propagation within the GCN architecture.

Conclusions:

• LaenNet offers a simple yet effective solution for improving GCN robustness in the presence of noisy and sparse graph data.
• The label propagation mechanism within LaenNet significantly enhances performance in challenging real-world graph scenarios.
• The LaenNet architecture is generalizable and provides a valuable contribution to the field of graph representation learning.