Hoang Van Thuy
PhD student
Network Science Lab
E-mail: hoangvanthuy90@gmail.com
WWW, May 2019, San Francisco, USA

1
Graph-structured Data
 Graph-structured Data
 Graph-structured data are ubiquitous.
 Graph-structured data are flexible to model complex interactions.
DBLP publication

2
Graph-structured Data
 Graph-structured Data
 Graph-structured data are ubiquitous.
 Graph-structured data are flexible to model complex interactions.
Protein-Proteins interactions

3
Graph Neural Network
 Neural networks for processing graph-structured inputs.
 Flexible to characterize non-Euclidean data.
 For example, graph convolutional network and graph attention network.

4
Heterogeneous Graphs
 Graph in real world
 Many node types, link types
 Non- ordered
 Complex connections

5
Heterogeneous Graphs
 Multiple types of nodes or links

6
Heterogeneous Graphs
 Multiple types of nodes or links
 Rich semantic information
 Meta-path: a relation sequence connecting objects
(e.g., Movie-Actor-Movie).

7
Meta-paths: High level representation of relationship
 DBLP Bibliographic network
 Node (type)
 KDD (Venue)
 Author 1
 Link (Type)
 Write ( Author - Paper)
 Publish ( Paper – Venue)

8
Motivation
 Existing Graph Neural Networks focus on homogeneous graph
 Cannot handle multiple types of nodes and edges.
 Cannot capture rich semantic information.

9
Challenges
 Challenges
 How to handle the heterogeneity of graph?
 How to discover the differences of meta-path based neighbors?
 How to find some meaningful meta-paths?

10
Overall Framework

11
Node-Level Attention and Aggregating
 X

12
Semantic-Level Attention and Aggregating
 X

14
Experiments
 Baselines:
 Deepwalk
 Esim
 Metapath2vec
 HERec
 GCN
 GAT
 HANnd
 HANsem
 Tasks:
 Node Classification
 Node Clustering
 Analysis of Attention Mechanism
 Visualization

15
Datasets

16
Node Classification
 X

17
Node Clustering & Visualization
 Visualization embedding on DBLP:
Each point: author, color: research area

18
Attention Analysis
 Node-Level Attention
 Given a meta-path PAP:
 P831,P699,P133: Data Mining
 P2384, P2328: Database
 P1973: Wireless Communication.
 Semantic-Level Attention

19
Conclusions
 The first attempt to study the heterogeneous graph neural network based on
attention mechanism.
 A novel heterogeneous graph attention network (HAN) which includes both of the
node-level and semantic-level attentions.
 The state-of-art performance and good interpretability

20