[Abstract]

• introduce G2GNN which alleviates the graph imbalance issue
• globally, construct GoG based on kernel similarity
• locally, topological augmentation via masking node features

[1. Introduction]

• real case: imblanace
  • ex) chemical space (active - inactive)
• problems of GNN on imbalanced datasets
  • incliniation to learning towards majority classes
  • poor generalization from gien scarce training data to abounding unseen testing data
• current solutions
  • augmenting data via under- or over- sampling
  • assigning weights
  • constructing synthetic training data

  ⇒ designed on point-based data

  ⇒ performance on graph-structured data is unclear

• graph-structured data
  • pre-training
  • adversarial training
• propose Graph-of-Graph Neural Networks (G2GNN)

[2. Problem formulation]

[3. Related work]

• graph imbalance problem
  • current learning works are for node imbalance classification.
  • graph imbalance classification remains largely unexplored
• graph of graphs
• graph augmentations

[4. The proposed framework]

• global governace: GoG propagation
• local explorer: topological augmentation

[4.1 Global Imbalance Mitigation: Graph-of-Graph Construction / Propagation]

• intuition: SMOTE, mixup (handling class imblance)

[4.1. Basic GNN encoder]

• GIN
• global-sum pooling

[4.1.2 Graph of Graphs Construction]

• edge of GoG: based on their topological similarity
  • leverage the graph kernel (Shortest Path Kernel)

• edge homophily

  

  

[4.1.3 Graph of Graphs Propagation]

• l-th-layer GoG propagation

  

[4.2 Local Imbalance Mitigation: Self-consistency Regularization via Graph Augmentation]

• augmentation

[4.2.1 Removing edges]

[4.2.2 Masking node features]

• zeroing entire features of some nodes

[4.2.3 Self-Consistency Regularization]

[4.3 Objective Function and Prediction]

[4.4 Algorithm]