NS-CUK Seminar: J.H.Lee, Review on "Learnable Structural Semantic Readout for Graph Classification", ICDM 2021

1. Joo-Ho Lee
School of Computer Science and Information Engineering,
The Catholic University of Korea
E-mail: jooho414@gmail.com
2023-08-21

2. 1
Introduction
Problem Statement
• Graph classification refers to the task of predicting class labels of input graphs, and it has been applied to a
wide range of graphs, including molecular structures, biological networks, and social networks
• The key challenge is to extract informative (or discriminative) graph features from topological structures (i.e.,
nodes and edges) and auxiliary node features

3. 2
Introduction
Problem Statement
• A recent challenge for improving the expressivity of the GNNs is leveraging the global (or graph-level)
structural information
• Most studies have developed GNN modules that capture such global position information into the
representations at the node-level or at the subgraph-level
 structure-aware message passing which further considers structural information of neighbor
nodes
 graph pooling based on spectral clustering

4. 3
Introduction
Problem Statement
• Despite these efforts, the global structural information has not been carefully considered at the graph-level
representations yet
• Note that every GNN classifier uses a global readout operation, which simply aggregates all remaining node
(or subgraph) representations, to obtain a permutation-invariant graph-level representation
• In this work, we point out that the global readout does not consider the structural information of each node,
which incurs information loss on the global structure of an input graph

5. 4
Introduction
Problem Statement
• To tackle this challenge, we propose a novel graph readout technique, Structural Semantic Readout (SSRead),
that outputs the graph-level representation explicitly keeping the global structural information
• Motivated by consistently-morphed graphs in the latent space according to its structural semantic, our
SSRead takes advantage of consistent positions in the latent space, which eventually correspond to the
structurally-meaningful positions in each graph

6. 5
Introduction
Contribution
• Compatibility
 SSRead can be easily embedded into any GNN architectures (i.e., compatible with a variety of message
passing and graph pooling layers) and make use of any aggregation functions for its position-level
readout
• Performance
 SSRead improves the classification accuracy by learning the position-level graph representations with
the position-aware classification layer, which exploits the global structural information
• Interpretability
 With the structural prototypes optimized from training data, SSRead can perform segmentation on a
graph according to the structural semantic and localize further discriminative regions for the target class

7. 6
Methodology
GNN Classifier

8. 7
Methodology
SSRead

9. 8
STRUCTURAL SEMANTIC READOUT
Semantic alignment with structural prototypes
• structural positions
ℳ ⊂ 0, 1 𝑁×𝐾

10. 9
TERACON
Position-level readout based on node-position alignment
• a summarized vector for structural position k
• “SSRead is a permutation-invariant function”

11. 10
TERACON
soft-SSA
• soft-min instead of hard-min

12. 11
TERACON
SSRead
• Alogorithm

13. 12
Experiment
Datasets

14. 13
Experiment
Architecture of GNN classifiers

15. 14
Experiment
Classification accuracy of the base GNN classifier

16. 15
Experiment
Classification accuracy of the state-of-the-art GNN classifiers

17. 16
Experiment
Classification accuracy on the graph-level representations

18. 17
Experiment
Performance of GRead and SSRead in predicting global structural properties

19. 18
Experiment
Visualization of our structural semantic alignment

20. 19
Experiment
Performance change varying the number of structural positions

21. 20
Conclusion
• This paper proposes a novel graph readout technique, named as SSRead, which outputs structured (or
position-level) representations in order to explicitly leverage the global structural information for graph
classification
• To this end, SSRead first identifies the structural position of the nodes by using the semantic alignment
between the node representations and the structural prototypes, which are optimized to best summarize K
structural semantics observed in the training graphs
• Their experiments support that SSRead consistently enhances the classification performance and
interpretability of the GNN classifier while providing great compatibility with various aggregation functions,
GNN architectures, and learning frameworks