32
Peng Cui Tsinghua University Perspectives and Outlook on Network Embedding and GCN

Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

  • Upload
    others

  • View
    18

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Peng Cui

Tsinghua University

Perspectives and Outlook on

Network Embedding and GCN

Page 2: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

2

The general description of data and their relations.

Network (Graph)

Page 3: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

3

Many types of data are networks

Social Networks Biology Networks Finance Networks

Internet of Things Information Networks Logistic Networks

Page 4: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

4

Why network is important?

In few cases, you only care about a subject but not its

relations with other subjects.

Reflected by relational subjects Decided by relational subjects

Target

Target

Image Characterization Social Capital

Page 5: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

5

G = ( V, E )

Networks are not learning-friendly

Links Topology

Inapplicability of

ML methods

Network

Data

Feature

Extraction

Pattern

Discovery

Network

Applications

Pipeline for network analysis

Page 6: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

6

Learning from networks

Network

EmbeddingGCN

Page 7: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

7

G = ( V, E ) G = ( V )Vector Space

generate

embed

• Easy to parallel

• Can apply classical ML methods

Network Embedding

Page 8: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

8

Goal Support network inference in vector space

Reflect network

structure

Maintain network

properties

B

A C

Transitivity

The goal of network embedding

Transform network nodes into vectors that are fit for

off-the-shelf machine learning models.

Page 9: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

9

Graph Neural Networks

Basic idea: recursive definition of states

A simple example: PageRank

F. Scarselli, et al. The graph neural network model. IEEE TNN, 2009.

Design a learning mechanism on graph.

Page 10: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Main idea: pass messages between pairs of nodes & agglomerate

Stacking multiple layers like standard CNNs:

State-of-the-art results on node classification

10

Graph Convolutional Networks (GCN)

T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. ICLR, 2017.

Page 11: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

11

A brief history of GNNs

Page 12: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Network Embedding and GCN

12

Graph

Feature

Network

Embedding

GCN

Input

Task results

Model Output

Embedding

Task results

Feature

Topology to Vector

Fusion of Topology and Features

Unsupervised v.s. (Semi-)Supervised

Page 13: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Graph convolutional network v.s. Network embedding

13

• In some sense, they are different.

• Graphs exist in mathematics. (Data Structure)

• Mathematical structures used to model pairwise relations between

objects

• Networks exist in the real world. (Data)

• Social networks, logistic networks, biology networks, transaction

networks, etc.

• A network can be represented by a graph.

• A dataset that is not a network can also be represented by a graph.

Page 14: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

GCN for Natural Language Processing

• Many papers on BERT + GNN.

• BERT is for retrieval.

• It creates an initial graph of relevant

entities and the initial evidence.

• GNN is for reasoning.

• It collects evidence (i.e., old messages

on the entities) and arrive at new

conclusions (i.e., new messages on

the entities), by passing the messages

around and aggregating them.

14

Cognitive Graph for Multi-Hop Reading Comprehension at Scale. Ding et al., ACL 2019.

Dynamically Fused Graph Network for Multi-hop Reasoning. Xiao et al., ACL 2019.

Page 15: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

GCN for Computer Vision

• A popular trend in CV is to construct a graph during the learning process.

• To process multiple objects or parts in a scene, and to infer their relationships.

• Example: Scene graphs.

15

Scene Graph Generation by Iterative Message Passing. Xu et al., CVPR 2017.

Image Generation from Scene Graphs. Johnson et al., CVPR 2018.

Page 16: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

GCN for Symbolic Reasoning

16

• We can view the process of symbolic reasoning as a directed acyclic graph.

• Many recent efforts use GNNs to perform symbolic reasoning.

Learning by Abstraction: The Neural State Machine. Hudson & Manning, 2019.

Can Graph Neural Networks Help Logic Reasoning? Zhang et al., 2019.

Symbolic Graph Reasoning Meets Convolutions. Liang et al., NeurIPS 2018.

Page 17: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

GCN for Structural Equation Modeling

• Structural equation modeling, a form of causal modeling, tries to describe the

relationships between the variables as a directed acyclic graph (DAG).

• GNN can be used to represent a nonlinear structural equation and help find

the DAG, after treating the adjacency matrix as parameters.

17

DAG-GNN: DAG Structure Learning with Graph Neural Networks. Yu et al., ICML 2019.

Page 18: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Pipeline for (most) GCN works

18

Raw DataGraph

ConstructionGCN End task

Page 19: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

• Co-occurrence (neighborhood)

19

Network embedding: topology to vector

Page 20: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

• High-order proximities

20

Network embedding: topology to vector

Page 21: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

• Communities

21

Network embedding: topology to vector

Page 22: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

• Heterogeneous networks

22

Network embedding: topology to vector

Page 23: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Pipeline for (most) Network Embedding works

23

Network

Data

Network

Embedding

Downstream

ModelEnd task

Page 24: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Learning for Networks vs. Learning via Graphs

24

Learning for networks

Learn

ing V

ia G

raphs

Network

Embedding

GCN

Page 25: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

The intrinsic problems NE is solving

25

Reducing representation dimensionality while preserving necessary

topological structures and properties.

Nodes & Links

Node Neighborhood

Community

Pair-wise Proximity

Hyper Edges

Global Structure

Non-transitivity

Asymmetric Transitivity

Dynamic

Uncertainty

Heterogeneity

Interpretability

Page 26: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

The intrinsic problem GCN is solving

26

Fusing topology and features in the way of smoothing features with

the assistance of topology.

N

N

N

d

N

d

X =

Page 27: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

What if the problem is topology-driven?

27

Since GCN is filtering features, it is inevitably feature-driven

Structure only provides auxiliary information (e.g. for filtering/smoothing)

When feature plays the key role, GNN performs good …

How about the contrary?

Synthesis data: stochastic block model + random features

Method Results

Random 10.0

GCN 18.3±1.1

DeepWalk 99.0±0.1

Page 28: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Network Embedding v.s. GCN

28

Node

FeaturesTopology

Network

Embedding

Feature-based

LearningGCN

There is no better one, but there is more proper one.

Page 29: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Rethinking: Is GCN truly a Deep Learning method?

29

Recall GNN formulation:

𝐻 𝑘+1 = 𝜎 𝑆𝐻 𝑘 𝑊 𝑘 , 𝑆 = ෩𝐷−1/2 ሚ𝐴෩𝐷−1/2

How about removing the non-linear component:

𝐻 𝑘+1 = 𝑆𝐻 𝑘 𝑊 𝑘

Stacking multiple layers and add softmax classification:

𝑌 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 𝐻 𝐾

= 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 𝑆𝑆…𝑆𝐻 0 𝑊 0 𝑊 1 …𝑊 𝐾−1

= 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 𝑆𝐾𝐻 0 𝑊

Wu, Felix, et al. Simplifying graph convolutional networks. ICML, 2019.

High-order proximity

Page 30: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Rethinking: Is GCN truly a Deep Learning method?

30

This simplified GNN (SGC) shows remarkable results:

Node classification Text Classification

Wu, Felix, et al. Simplifying graph convolutional networks. ICML, 2019.

Page 31: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Summaries and Conclusions

31

Unsupervised v.s. (Semi-)Supervised

Learning for Networks v.s. Learning via Graphs

Topology-driven v.s. Feature-driven

Both GCN and NE need to treat the counterpart as the baselines

Page 32: Tsinghua University - Peng Cui (Cui, Peng)'s Homepagepengcui.thumedialab.com/papers/Perspectives on NE and GCN... · 2019. 8. 7. · Tsinghua University Perspectives and Outlook on

Thanks!

Peng Cui

[email protected]://pengcui.thumedialab.com

32