Hyperbolic Graph Contrastive Learning with Hierarchical Positives

Summary

在 hyperbolic space 中引入 contrastive learning framework，利用 hierarchical positive sampling 提升 graph representation learning。Node classification 和 graph property prediction 上达到 SOTA。

Problem & Motivation

Graph contrastive learning 问题：

• Euclidean contrastive learning 忽略 graph hierarchy
• Hierarchical data（知识图谱、社交网络）需要 geometry-aware sampling
• Positive/negative sample 定义需要 respect geometry

Method

核心设计：

1. Hierarchical Positive Sampling: 在 hyperbolic space 中按 hierarchy 采样 positives
2. Geometry-aware Contrastive Loss: 考虑 hyperbolic distance 的 loss design
3. Curvature-aware Training: Hyperbolic space 的 gradient handling

理论基础：

• Hyperbolic distance ≈ hierarchy distance
• Contrastive loss in curved space

Key Results

• Node classification SOTA
• Graph property prediction improved
• Few-shot learning scenarios

Strengths & Weaknesses

亮点：

• Hierarchical sampling 是关键创新
• Contrastive learning × hyperbolic geometry 的首次系统结合

局限：

• 具体 benchmark 数字需看全文
• 与 Euclidean contrastive learning 的效率对比

Mind Map

mindmap
  root((Hyperbolic Contrastive))
    Problem
      Euclidean contrastive limits
      Hierarchy ignored
    Method
      Hierarchical positives
      Geometry-aware loss
      Curvature-aware gradient
    Results
      Node classification SOTA
      Graph prediction improved

Notes

[基于 WebSearch 结果创建]

Contrastive learning 与 hyperbolic geometry 的结合是重要方向。Hierarchical positive sampling 的设计思路值得研究。