You are on page 1of 29

Poincaré Embeddings for

Learning Hierarchical
Representations
July 4, 2017
Tatsuya Shirakawa
ABEJA Inc.
CONFIDENTIAL

Tatsuya Shirakawa
CONFIDENTIAL

Today’s Paper

Paper Stats
• Guys from FAIR
• Sumitted to arXiv at May 26, 2017
https://arxiv.org/abs/1705.08039
• Sumitted to NIPS2017?

Key Contributions
• Introducing hyperbolic geometry
(Poincaré disk model) into word/graph
embeddings paradigm
• Automatically capture hierarchical
structure of data
• Achieved incredible better results than
previous works.
CONFIDENTIAL

Agenda

1. Problems
2. Hyperbolic Geometry
3. Poincaré Embeddings
(and Some Incredible Results)

4
CONFIDENTIAL

Problems

5
CONFIDENTIAL

The Problem

Find good representation(embedding) of items such


that underlying hierarchical relation structure are
well reconstructed
CONFIDENTIAL

Taxonomy Embedding

Embed nouns in WordNets such that related nouns


are close in embedded space

http://www.nltk.org/book_1ed/ch02.html
7
CONFIDENTIAL

Graph Link Prediction

Embed nodes in given graph such that missing


links are well-reconstructed

http://ml.cs.tsinghua.edu.cn/~jiaming/publications/

8
CONFIDENTIAL

Back Theory

9
CONFIDENTIAL

Hyperbolic Geometry

• Geometry with negative curvature


• Many models (realizations):
- Poincaré half space model
- Poincaré disk model

each is isometric

10
CONFIDENTIAL

Poincaré Half Space Model

• Defined on upper half space


with metric

• Distance btw points is

11
CONFIDENTIAL

Tree representation in H

• Tree structure is well


represented in Poincaré
half space

12
https://arxiv.org/abs/1006.5169
CONFIDENTIAL

Poincaré Disk Model

• A realization of hyperbolic geometry


• Defined on
equipped with metric of

• Distance btw points is

M.C. Escher's Circle Limit III, 1959


13
CONFIDENTIAL

Relation to Poincaré Half Space Model

(for simplicity: 2-dim, identify as )

https://arxiv.org/abs/1006.5169
14
CONFIDENTIAL

Why not Euclidean Space?

• Euclidean space is too narrow to embed


hierarchical (tree) structures

Surface Area Volume


/ # of leaf nodes / # of nodes

Euclidean Ball O(R^n) O(R^n)

b-ary tree O(b^R) O(b^R)

※ R=radius of ball or depth of tree

15
CONFIDENTIAL

Why Hyperbolic Space?

• b-array tree can be interpreted as discrete


analogue of Poincaré disk

16
CONFIDENTIAL

Conclusion Here

• Hyperbolic space is far more appropriate than


Euclidean space to represent hierarchical
structure

• Many equivalent models


- Poincaré half space model
- Poincaré disk model

17
CONFIDENTIAL

Prior Works around hyperbolic geometry


applications
• R. Kleinberg, “Geographic routing using hyperbolic
spaces”, 2007
• M. Boguna et al., “Sustaining the internet with
hyperbolic mapping”, 2010
• P. D. Hoff et al., “Latent space approaches to social
network analysis”, 2016
• A. B. Adcock et al., “Tree-like structure in large social
and information networks’, 2013
• D. Krioukov et al., “Hyperbolic geometry of complex
networks”, 2010

18
CONFIDENTIAL

Poincaré Embeddings

19
CONFIDENTIAL

Proposed Method

1. Parametrize each item in Poincaré ball

2. Optimize them by Riemannian optimization


under metric of
CONFIDENTIAL

Riemannian SGD

1. Compute stochastic (Euclidean) gradient

2. Correct metric

3. Apply GD

4. Project onto space

21
CONFIDENTIAL

Taxonomy Embedding

Embed nouns in WordNets such that related nouns


are close in embedded space

http://www.nltk.org/book_1ed/ch02.html
22
CONFIDENTIAL

Objective Function

Maximize

randomly chosen 10 negative samples

Reconstruction setting:
- D is full relations
Prediction setting
- D is subset of full relations

23
CONFIDENTIAL

Result

24
CONFIDENTIAL

25
CONFIDENTIAL

Graph Link Prediction

Embed nodes in given graph such that missing


links are well-reconstructed

http://ml.cs.tsinghua.edu.cn/~jiaming/publications/

26
CONFIDENTIAL

Objective Function

Minimize the cross entropy of probability

27
CONFIDENTIAL

Result

28
CONFIDENTIAL

Summary

• Poincaré embeddings automatically capture


hierarchical structure from data

• Riemannian SGD provides the way to optimize


Poincaré embeddings

• Achieved quite good results on word/graph


embedding tasks

29

You might also like