DigGCN: Learning Compact Graph Convolutional Networks via Diffusion Aggregation.

Area of Science:

• Machine Learning
• Network Science
• Artificial Intelligence

Background:

• Graph neural networks (GNNs) excel at network embedding via message passing.
• Current GNNs struggle with static structures, failing to capture latent/high-order information and facing over-smoothing issues.
• Existing aggregation methods limit receptive fields to visible connections.

Purpose of the Study:

• To propose a compact graph convolutional network (GCN) framework addressing GNN limitations.
• To learn from invisible connections and recover latent proximity in networks.
• To develop a lightweight GCN model that avoids overfitting while utilizing deep information.

Main Methods:

• Inferred high-order proximity and constructed diffusion paths using diffusion samplings (region-based).
• Employed network inference to obtain accurate weights for building informative receptive fields with salient neighbors.
• Introduced a nonconvex regularizer to learn a lightweight model, mitigating overfitting and depth dilemma.

Main Results:

• The proposed GCN framework effectively learns from invisible connections and recovers latent proximity.
• Diffusion samplings proved more effective than random walk samplings for constructing diffusion paths.
• The model demonstrated effectiveness in unsupervised feature learning and supervised classification tasks.

Conclusions:

• The compact GCN framework successfully overcomes limitations of traditional GNNs regarding static structures and over-smoothing.
• The approach enables learning from latent structures and enhances network embedding capabilities.
• The method offers a promising solution for efficient and effective graph representation learning.