NS-CUK Seminar: J.H.Lee, Review on "MGCN: Semi-supervised Classification in Multi-layer Graphs with Graph Convolutional Networks", ASONAM 2019

1. Joo-Ho Lee
School of Computer Science and
Information Engineering, The Catholic
University of Korea
E-mail: jooho414@gmail.com
2023-04-07

2. 1
 Introduction
• Problem statement
• Contributions
 Related Works
 Methodology
 Experiment
 Conclusion

3. 2
Introduction
Problem Statement
• Most of the existing methods only focus on single-layer graphs
• However, in many real-world tasks, nodes and/or edges of the graphs have different types
• Multi-layer graphs provide a framework to accommodate different types of entities and relations

4. 3
Introduction
Contribution
• It provides a framework for node embedding of multilayer graphs using GCNs for the first time to the best of our
knowledge
• Due to the deep architecture, our method can be trained end-to-end as opposed to the most of the existing
multilayer graph embedding methods
• The proposed loss function is composed of the structure reconstruction error and the classification error. It is
also able to embed node attributes simultaneously

5. 4
Methodology
Architecture

6. 5
Methodology
Unsupervised Learning
• Unsupervised Learning : Structural Information in intra layer and inter layer
• Supervised Learning : Characteristic of node in graph

7. 6
Methodology
Unsupervised Learning
 This mean reconstruct

8. 7
Methodology
Supervised Learning

9. 8
Experiment
Datasets
• Infra (Infrastructure system network)
• Aminer (paper – people – conference)
Baselines
• MANE
• DeepWalk
• Node2Vec
• GCN

10. 9
Experiment

11. 10
Experiment

12. 11
Conclusion
• In this paper, we extended the GCN model to embed multi-layer graph structure (and also node attributes when
available)
• And they extended the GCN model to propose an end-to-end deep learning method, named MGCN, that is able
to find representation of nodes considering all available information for semi-supervised classification
• This paper showed the superiority of MGCN in considering between-layer edges to some single-layer graph
embedding methods and also to a recent multi-layer graph embedding method on the node classification task