Scalable Hyperbolic Representation Learning for Large-Scale Knowledge Graphs

Summary

大规模知识图谱的 scalable hyperbolic embedding，通过 efficient curvature learning 和 mini-batch training 支持百万实体规模的 graph embedding。

Problem & Motivation

Knowledge graph embedding scalability 问题：

• Hyperbolic embedding 计算成本高
• Million-entity graphs 无法端到端训练
• Curvature learning 在大规模数据上不稳定

Method

核心设计：

1. Mini-batch Training: Hyperbolic space 的 mini-batch gradient descent
2. Efficient Curvature Learning: Data-adaptive curvature estimation
3. Scalable Riemannian Optimization: 大规模 RSGD 实现

技术挑战：

• Tangent space ↔ manifold mapping cost
• Mini-batch manifold operations

Key Results

• 百万实体 KG embedding feasible
• Curvature learning stable
• Link prediction improved

Strengths & Weaknesses

亮点：

• Scalability 是 hyperbolic embedding 的核心痛点
• Mini-batch training 是重要工程突破

局限：

• Efficiency vs accuracy trade-off？
• 与 Euclidean KG embedding 的效率对比

Mind Map

mindmap
  root((Scalable Hyperbolic KG))
    Problem
      Million entities
      Training cost high
      Curvature unstable
    Method
      Mini-batch training
      Efficient curvature learning
      Scalable RSGD
    Results
      Million entities feasible
      Link prediction improved

Notes

[基于 WebSearch 结果创建]

Scalability 是 hyperbolic embedding 的核心瓶颈。Mini-batch training 和 efficient curvature learning 是关键工程贡献。