๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
2๏ธโƒฃ Study/๐Ÿ“‘ ๋…ผ๋ฌธ์ฝ๊ธฐ

Graph Clustering with Graph Neural Networks (2020)

by isdawell 2022. 12. 23.
728x90

 

 

๋…ผ๋ฌธ 

 

gnn clustering.pdf
2.06MB

 

 

 

 

 

1๏ธโƒฃ  Abstract 


 

 

 

 

โ‘  ๋ฌธ์ œ์  ์ œ์‹œ 

 

• graph clustering ์—์„œ "๋น„์ง€๋„ํ•™์Šต" ์— ๊ด€ํ•œ ๋ฐœ์ „์ด ๋”๋””๊ฒŒ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค. 

• graph clustering ๊ธฐ๋ฒ•์ด ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค. 

 

 

โ‘ก ํ•ด๊ฒฐ ๋ฐฉ์•ˆ 

 

DMoN : Deep Modularity Networks : ๋น„์ง€๋„ ํ’€๋ง ํ•™์Šต ๋ฐฉ๋ฒ• 

 

โ–ธ graph pooling : https://greeksharifa.github.io/machine_learning/2021/09/09/GraphPooling/

 

Python, Machine & Deep Learning

Python, Machine Learning & Deep Learning

greeksharifa.github.io

 

โ†ช ๋…ธ๋“œ ์ž„๋ฒ ๋”ฉ๋“ค๋กœ๋ถ€ํ„ฐ ๊ทธ๋ž˜ํ”„ ์ž„๋ฒ ๋”ฉ์„ ์–ป๋Š” ๊ณผ์ •์„ ์˜๋ฏธํ•œ๋‹ค. ๊ฐ„๋‹จํ•˜๊ฒŒ ๋งํ•˜์ž๋ฉด, full graph, ์ฆ‰ ์ „์ฒด ๊ทธ๋ž˜ํ”„์—์„œ ํฌ๊ธฐ์™€ ๋ฐ€๋„๋ฅผ ์ค„์—ฌ subgraph ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ณผ์ •์œผ๋กœ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค. ์ด๋•Œ subgraph ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ฒด ๊ทธ๋ž˜ํ”„์—์„œ ์ผ๋ถ€๋ถ„์„ ๋œปํ•จ๊ณผ ํ—ท๊ฐˆ๋ฆฌ์ง€ ์•Š์•„์•ผ ํ•œ๋‹ค. 

 

 

โ–ธ Modularity Clustering : ๊ทธ๋ž˜ํ”„๋ฅผ ๊ณ„์ธต์ ์œผ๋กœ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ, ํด๋ž˜์Šคํ„ฐ ๋‚ด๋ถ€์˜ ๋งํฌ์™€ ๊ทธ ๋งํฌ๋“ค๊ฐ„์˜ ๋ฐ€์ง‘๋„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ modularity ๋ฅผ ๊ณ„์‚ฐํ•ด ๊ฐ€์žฅ ๋†’์€ modularity ๋ฅผ ๊ฐ–๋Š” ํด๋Ÿฌ์Šคํ„ฐ๋“ค๋กœ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ• 

 

โ†ช modularity : measures the strength of division of a network into modules (also called groups, clusters or communities). Networks with high modularity have dense connections between the nodes within modules but sparse connections between nodes in different modules.

 

 

 

2๏ธโƒฃ  Introduction 


 

โ‘   Cluster in Graph 

 

• ๊ทธ๋ž˜ํ”„๋Š” cluster (= communities = modules) ๋“ค์„ ํ˜•์„ฑํ•˜๋ฉฐ ์ง€์—ญ์ ์œผ๋กœ ๋น„๋™์งˆ์ (inhomogeneous) edge ๋ถ„ํฌ๋ฅผ ๋ณด์ธ๋‹ค. 

• cluster ๋Š” ๋†’์€ edge (link) ๋ฐ€์ง‘๋„๋ฅผ ๋ณด์ด๋Š” ๋…ธ๋“œ๋“ค์˜ ์ง‘ํ•ฉ์„ ์˜๋ฏธํ•œ๋‹ค. 

• cluster ๋ฅผ ์ •์˜ํ•จ์œผ๋กœ์จ ๊ต์œก์ด๋‚˜ ๊ณ ์šฉ ๋ถ„์•ผ์™€ ๊ฐ™์€ social graph ์—์„œ ํฅ๋ฏธ๋กœ์šด ํ˜„์ƒ๋“ค์„ ๋ฐœ๊ฒฌํ•ด๋‚ผ ์ˆ˜ ์žˆ๋‹ค. 

 

 

โ‘ก  Problem of Recent works about GNNs (in Clustering)

 

• ๊ณ ์ฐจ์› ๊ตฌ์กฐ๋ฅผ ํ™œ์šฉํ•˜๋Š” GNN ์˜ ๋Œ€๋ถ€๋ถ„์— ๊ด€ํ•œ ์—ฐ๊ตฌ๋Š” node partitioning ์ด๋‚˜ clustering ์— ์ง์ ‘์ ์œผ๋กœ ๊ด€์—ฌํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋“œ๋ฌผ๋‹ค. ํŠนํžˆ ๋น„์ง€๋„ํ•™์Šต์˜ ๊ทธ๋ž˜ํ”„ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๋ฐฉ๋ฒ•์„ ์ œ์™ธํ•˜๊ณ  semi-supservised ํ˜น์€ supervised ํ”„๋ ˆ์ž„์›Œํฌ๋งŒ์„ ๋‹ค๋ฃจ๋Š” ์‚ฌ๋ก€๊ฐ€ ๋งŽ๋‹ค. 

 

โ€ป semi-supervised learning : ์†Œ๋Ÿ‰์˜ labeled data์—๋Š” supervised learning์„ ์ ์šฉํ•˜๊ณ  ๋Œ€์šฉ๋Ÿ‰ unlabeled data์—๋Š” unsupervised learning์„ ์ ์šฉํ•ด ์ถ”๊ฐ€์ ์ธ ์„ฑ๋Šฅํ–ฅ์ƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก . unlabeled data ๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— generalization ์— ์œ ๋ฆฌํ•˜๋‹ค. 

 

• ํ˜„์กดํ•˜๋Š” ๋น„์ง€๋„ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋Œ€๋ถ€๋ถ„์€ multi-step optimization ๋ฐฉ์‹์— ์˜์กดํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ข…๋‹จ์˜ ๋ฏธ๋ถ„๊ฐ€๋Šฅํ•œ ๋ชฉ์ ํ•จ์ˆ˜ (end-to-end differentiable objective) ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ๋ฌธ์ œ์ ์ด ์žˆ๋‹ค. 

 

 

โ‘ข Solution 

 

• Conntection between Graph Pooling ( supervised GNN ) and fully unsupervised Clustering ๐Ÿ‘‰ DMoN

 

โ†ช GNN ์„ ํ™œ์šฉํ•œ ๋น„์ง€๋„ํ•™์Šต ํด๋Ÿฌ์Šคํ„ฐ๋ง ๋ชจ๋“ˆ์ธ DMoN ์€ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ณผ์ œ์—์„œ ์ข…๋‹จ์˜ ๋ฏธ๋ถ„๊ฐ€๋Šฅํ•œ ๋ฐฉ์‹์œผ๋กœ  ์ตœ์ ํ™” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•œ๋‹ค. 

 

 

 

3๏ธโƒฃ  Related Works 


 

โ‘   GNNs

 

• allow end-to-end differentable losses over data with arbiturary structure 

• applications : social networks, recommender systems, computational chemistry 

• GNN ์€ ๋น„์ง€๋„ํ•™์Šต์˜ ์†์‹คํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ด๋„ ๋  ๋งŒํผ ์œ ์—ฐํ•œ ๋„คํŠธ์›Œํฌ์— ํ•ด๋‹นํ•˜์ง€๋งŒ, node classification ๋“ฑ์˜ ๋Œ€๋ถ€๋ถ„์˜ ์—ฐ๊ตฌ๋Š” semi-supervised ํ™˜๊ฒฝ์—์„œ task ๋“ค์„ ๋‹ค๋ฃจ๊ณ  ์žˆ๋‹ค. 

 

 

โ‘ก  Unsupervised training of GNNs 

 

• GNN ์˜ ๋น„์ง€๋„ํ•™์Šต์€ ๋ณดํ†ต mutual information ์„ ์ตœ๋Œ€ํ™” ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ํ•™์Šต๋œ๋‹ค. 

 

โ€ป mutual information : ํ•˜๋‚˜์˜ random variable ์ด ๋‹ค๋ฅธ random variable ์— ์–ผ๋งˆ๋‚˜ dependent ํ•œ์ง€๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ง€ํ‘œ๋กœ P(X,Y) ๊ฐ€ P(X)P(Y) ์™€ ์–ผ๋งˆ๋‚˜ ๋น„์Šทํ•œ์ง€๋ฅผ ์ธก์ •ํ•œ๋‹ค. independent ๊ด€๊ณ„๋ผ๋ฉด ๋‘ ๊ฐ’์ด ๊ฐ™์œผ๋ฉฐ MI ๊ฐ’์€ 0์ด ๋œ๋‹ค. 

 

• ๊ด€๋ จ ์—ฐ๊ตฌ : Deep Graph Informax(2019), InfoGraph (2020) 

 

 

โ‘ข  Graph Pooling 

 

• Graph Pooling ์€ ๋ฐ˜๋ณต์ ์ธ data coarsening (๋ฒ”์ฃผ๋ฅผ ๊ฐ„๋‹จํ™” ์‹œํ‚ค๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ฐ์ดํ„ฐ๋ฅผ ํ•จ์ถ•์‹œํ‚ค๋Š” ๊ฒƒ์ด๋ผ ์ƒ๊ฐํ•˜๋ฉด ๋จ) ์„ ํ†ตํ•ด ๊ทธ๋ž˜ํ”„์˜ ๊ณ„์ธต์  ํŠน์ง•์„ ๋ฐœ๊ฒฌํ•ด๋‚ด๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ 

 

• ์ดˆ๊ธฐ์˜ ๊ตฌ์กฐ๋Š” ํด๋Ÿฌ์Šคํ„ฐ๋ง ์ตœ์ ํ™” ๊ณผ์ •์ด ์—†๋Š” ๊ณ ์ •๋œ axiomatic pooling ๋ฐฉ์‹์— ์˜์กดํ•˜๋‹ค๊ฐ€ DiffPool (2018) ์„ ์ œ์•ˆํ•œ ๋…ผ๋ฌธ์—์„œ ์ฒ˜์Œ์œผ๋กœ GNN ๊ตฌ์กฐ์—์„œ ํ•™์Šต๊ฐ€๋Šฅํ•œ ํ’€๋ง์„ ์ œ์•ˆํ•œ ๋ฐ”๊ฐ€ ์žˆ๋‹ค. 

 

• MinCutPool (2020) ์„ ์ œ์•ˆํ•œ ์—ฐ๊ตฌ์—์„œ๋Š” ํ’€๋ง ์ „๋žต์œผ๋กœ์„œ spectral clustering ์˜ ๋ฏธ๋ถ„๊ฐ€๋Šฅํ•œ ๊ณต์‹์— ๋Œ€ํ•ด ์†Œ๊ฐœํ•œ ๋ฐ”๊ฐ€ ์žˆ๋‹ค. MinCutPool ์€ ๋ณต์žกํ•œ ๊ทธ๋ž˜ํ”„์—์„œ ์ฃผ์š”ํ•œ ๋…ธ๋“œ๋ฅผ ์ถ”์ถœํ•˜๊ณ ์ž ํ•œ ์—ฌ๋Ÿฌ ์‹œ๋„๋“ค ์ค‘ ํ•˜๋‚˜๋กœ eigenvector ๋ฅผ ํ™œ์šฉํ•ด์•ผํ•˜๋Š” spectral clustering ์˜ ๋‹จ์ ์„ ํ•ด์†Œํ•˜๋ฉด์„œ node feature ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•œ๋‹ค. 

 

โ€ป Spectral clustering : ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜์˜ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ธฐ๋ฒ•์œผ๋กœ, k-means clustering ๊ณผ ๊ฐ™์ด ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ๊ฑฐ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฐ์ง‘ํ™”๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ, ๊ฐ ๋…ธ๋“œ๋ณ„๋กœ ์—ฐ๊ฒฐ์ด ๋˜์–ด์žˆ๋Š”์ง€, ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋‹ค๋ฉด ์–ผ๋งˆ๋งŒํผ์˜ ๊ฐ•๋„๋กœ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋Š”์ง€๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ๊ตฐ์ง‘ํ™”๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ํ•ด๋‹นํ•œ๋‹ค. 

 

๐Ÿ‘€ ์ฐธ๊ณ  : https://velog.io/@tobigs-gnn1213/5.-Spectral-Clustering#1-spectral-clustering-algorithms

 

5. Spectral Clustering

Spectral Clustering [์ž‘์„ฑ์ž : ์ •๋ฏผ์ค€]

velog.io

 

๐Ÿ‘€ ์ฐธ๊ณ  : https://seungwooham.github.io/machine%20learning/MinCutPool1/

 

MinCutPool: ์ด๋ก ํŽธ1 (MinCutPool: Understanding the Theory 1)

MinCutPool: ์ด๋ก ํŽธ1 (MinCutPool: Understanding the Theory 1)

seungwooham.github.io

 

 

๐Ÿ’ก Graph pooling ์— ๊ด€ํ•œ properties 

 

End-to-End training ์ข…๋‹จํ•™์Šต์€ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ์™€ ๋…ธ๋“œ ํŠน์„ฑ์„ ๋™์‹œ์— ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•œ๋‹ค. 
Unsupervised ๋น„์ง€๋„ํ•™์Šต์€ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๋ชจ๋ธ๋ง์— ์žˆ์–ด desirable ํ•œ ํ•™์Šต๋ฐฉ๋ฒ•์ด๋‹ค. 
Node aggregation ํด๋Ÿฌ์Šคํ„ฐ๋ง์—์„œ graph pooling ์„ ํ•ด์„ํ•˜๋Š”๋ฐ ์žˆ์–ด ์ค‘์š”ํ•œ ๋ถ€๋ถ„์ด๋‹ค. 
Sparse  ์‹ค์ œ ๊ทธ๋ž˜ํ”„ ๋ฐ์ดํ„ฐ๋Š” size ๋„ ๋‹ค์–‘ํ•˜๋ฉฐ sprsity ๊ฐ€ ์กด์žฌํ•œ๋‹ค. 
Soft aassignment  cluster ๋ฅผ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์— ์žˆ์–ด soft ํ•œ ๋ฐฉ์‹์„ ๋„์ž…ํ•˜๋ฉด ํด๋Ÿฌ์Šคํ„ฐ ๊ฐ„ ์ƒํ˜ธ์ž‘์šฉ์„ ํ•ด์„ํ•˜๋Š”๋ฐ์—๋„ ์œ ์—ฐํ•œ ์ถ”๋ก ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. 
Stable  ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ์— ๊ด€ํ•˜์—ฌ, ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ์ง„ํ–‰ํ•  ๋•Œ ์‚ฌ์šฉํ•˜๋Š” ํ•จ์ˆ˜๋Š” stable ํ•ด์•ผ ํ•œ๋‹ค. 

 

โ€ป DMoN ์€ subsampling ๋กœ ์ธํ•œ ์ •๋ณด์˜ ์†์‹ค ์—†์ด, Graph scalability ์ธก๋ฉด์—์„œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค. 

 

 

 

 

4๏ธโƒฃ  Preliminaries 


 

โ‘   Notation 

 

• V : set of nodes 

• E : set of edges 

• A : nxn adjacency matrix of G 

• di : degree of vi 

• F : graph partitioning function → V ๋ฅผ k ๊ฐœ์˜ ์ง‘๋‹จ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ํ•จ์ˆ˜  (for data clustering) 

• X : node attributes 

 

 

 

โ‘ก  Graph Clustering Quality Functions 

 

• DMoN ์€ spectral ์ ‘๊ทผ์„ ์‚ฌ์šฉํ•ด gradient ๊ธฐ๋ฐ˜์˜ ์ตœ์ ํ™” ๋ฐฉ์‹์— ์ ํ•ฉํ•œ ๋ชฉ์ ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. 

• spectral Clustering : Graph adjacency matrix ์˜ eigenvalue, eigenvector ๋ฅผ ์‚ฌ์šฉํ•ด ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉ์‹ 

 

(1) Cut based metrics 

 

๐Ÿ‘€ ์ฐธ๊ณ  : https://techblog-history-younghunjo1.tistory.com/93

 

[ML] Spectral Clustering(์ŠคํŽ™ํŠธ๋Ÿด ํด๋Ÿฌ์Šคํ„ฐ๋ง)

์ด๋ฒˆ ํฌ์ŠคํŒ…์—์„œ๋Š” ํด๋Ÿฌ์Šคํ„ฐ๋ง ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜๋กœ์„œ ์ŠคํŽ™ํŠธ๋Ÿด ํด๋Ÿฌ์Šคํ„ฐ๋ง์— ๋Œ€ํ•ด์„œ ์†Œ๊ฐœํ•˜๋ ค ํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ ์ŠคํŽ™ํŠธ๋Ÿด(Spectral) ์ž์ฒด๋Š” ํ–‰๋ ฌ์˜ ๊ณ ์œณ๊ฐ’(Eigenvalue)์„ ์˜๋ฏธํ•œ๋‹ค. ์ด๋Š” ์ถ”ํ›„์— ๊ณ ์œณ๊ฐ’์„ ์–ด

techblog-history-younghunjo1.tistory.com

 

๐Ÿ‘€ ๋‘ ๊ทธ๋ฃน์œผ๋กœ ๋‚˜๋ˆ„๋Š” (Cut) ๋ฐฉ์‹ 2๊ฐ€์ง€ 

 

cut minimal (minimum cut)  : ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ๋“ค ์‚ฌ์ด์˜ ์œ ์‚ฌ๋„ ๊ฐ’๋“ค ์ค‘ ๊ฐ€์žฅ ์ž‘์€ ๊ฐ’์œผ๋กœ ๊ฐ๊ฐ ํด๋Ÿฌ์Šคํ„ฐ๋กœ ๋‚˜๋ˆˆ๋‹ค. ์ฒซ๋ฒˆ์งธ ๊ณ ์œ ๊ฐ’์€ ๋ณดํ†ต ๋ฌด์˜๋ฏธํ•œ ํ•ด์ด๊ธฐ ๋•Œ๋ฌธ์— ๋‘๋ฒˆ์งธ ๊ณ ์œณ๊ฐ’๋ถ€ํ„ฐ ์‚ฌ์šฉํ•œ๋‹ค. (second eigenvector ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋น„์Šทํ•œ ๋งฅ๋ฝ) 

 

minimum cut ์˜ Unoptimal ํ•œ ๊ฒฝ์šฐ

 

ratio cut (normalized cut) : minimal cut ๋ฐฉ๋ฒ•์€ ๊ฐ€์ค‘์น˜ ๊ฐ’์ด ์ž‘์€ ์—ฃ์ง€๊ฐ€ ํด๋Ÿฌ์Šคํ„ฐ ๋‚ด๋ถ€์— ์กด์žฌํ•˜๊ฒŒ ๋œ๋‹ค๋ฉด ๋ถ„๋ฅ˜๊ฐ€ ์ž˜๋ชป ์ด๋ฃจ์–ด์ง„๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ partition ์˜ total edge volume ์„ ์ด์šฉํ•ด normalized cut ๋ฐฉ์‹์„ ๋„์ž…ํ•œ๋‹ค. ๋ณด๋‹ค ๊ท ํ˜•์ ์ธ partition ์„ ์œ„ํ•ด MinCutPool ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ๋„ normalized cut ๋ฐฉ์‹์„ ๋„์ž…ํ•œ๋‹ค. 

 

cut ์ด๋ž€ ๋‘ ๊ฐœ์˜ group ์„ ์–ป๊ธฐ ์œ„ํ•ด ๋ช‡๊ฐœ์˜ edge ๋ฅผ ๋Š์–ด์•ผ ํ•˜๋Š”๊ฐ€์— ๊ด€ํ•œ ๊ฐ’์ด๋‹ค.

 

 

(2) Modularity 

 

 Modularity ๋ชฉ์ ํ•จ์ˆ˜๋Š” "ํ†ต๊ณ„ํ•™์ " ๊ด€์ ์œผ๋กœ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๋ฌธ์ œ๋ฅผ ์ ‘๊ทผํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ, ๋„คํŠธ์›Œํฌ๊ฐ€ community ๋กœ ์–ผ๋งˆ๋‚˜ ์ž˜ ๋‚˜๋ˆ„์–ด์กŒ๋Š”๊ฐ€์— ๋Œ€ํ•œ ์ˆ˜์น˜๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๋žœ๋คํ•œ ๊ทธ๋ž˜ํ”„๋ฅผ ํ†ตํ•ด ์–ป์–ด์ง„ expectation ๊ฐ’์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ˆ˜์น˜๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” null model ์„ ๋„์ž…ํ•˜์—ฌ ํด๋Ÿฌ์Šคํ„ฐ๋ง์˜ ํŽธ์ฐจ๋ฅผ ์ˆ˜๋Ÿ‰ํ™” ํ•œ๋‹ค. 

 

 

๐Ÿ”น Modularity measure

 

โ†ช Fully Random Graph with given degrees, node u and v with degrees d(u) and d(v) are connected with probability d(u)โ—ฆd(v)/2m 

โ†ช n is # of nodes and m is # of edges 

 

Undirected Graph ๋ผ๊ณ  ๊ฐ€์ •ํ–ˆ์„ ๋•Œ, ๋ชจ๋“  ์—ฃ์ง€๋Š” 2๋ฒˆ์”ฉ ์„ธ์–ด์งˆ ์ˆ˜ ์žˆ๊ณ , ๋”ฐ๋ผ์„œ ์ด๋•Œ์˜ d(v)/2m ์€ ํŠน์ • ๋…ธ๋“œ๊ฐ€ ๋…ธ๋“œ v ์™€ ์—ฐ๊ฒฐ๋  ํ™•๋ฅ ๋กœ ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ (๊ฐ€๋Šฅํ•œ ์ „์ฒด ์—ฃ์ง€์˜ ๊ฐœ์ˆ˜ ๋ถ„์— v ์— ์—ฐ๊ฒฐ๋œ ์—ฃ์ง€์˜ ๊ฐœ์ˆ˜), multi-graph (๋‘ ๋…ธ๋“œ ์‚ฌ์ด์— ์—ฃ์ง€์˜ ๊ฐœ์ˆ˜๊ฐ€ ์—ฌ๋Ÿฌ ๊ฐœ ํ˜•์„ฑ๋  ์ˆ˜ ์žˆ๋Š” ๊ทธ๋ž˜ํ”„) ์ด๋ฏ€๋กœ ๋…ธ๋“œ u ์˜ ์ฐจ์ˆ˜๋ฅผ ๊ณฑํ•ด์ฃผ๋ฉด ๋…ธ๋“œ u ์™€ v ์‚ฌ์ด์— ๊ธฐ๋Œ€ํ•  ์ˆ˜ ์žˆ๋Š” ์—ฃ์ง€์˜ ๊ฐœ์ˆ˜๊ฐ€ ๋„์ถœ๋˜๋Š” ๊ฒƒ์ด๋‹ค. 

 

Modularity ๋Š” ์‹ค์ œ ๋‘ ๋…ธ๋“œ ์‚ฌ์ด์˜ ์—ฃ์ง€ ๊ฐœ์ˆ˜๋ฅผ ์˜๋ฏธํ•˜๋Š” Aij ์™€ ๊ทธ ๊ธฐ๋Œ“๊ฐ’์˜ ์ฐจ์ด๋ฅผ ๊ณ„์‚ฐํ•จ์œผ๋กœ์จ ๊ฒฐ๊ณผ๊ฐ€ ์–ป์–ด์ง€๋ฉฐ, Normalizing ์„ ์œ„ํ•ด 2m ์„ ๋‚˜๋ˆ„์–ด์คŒ์œผ๋กœ์จ -1๊ณผ 1์‚ฌ์ด์˜ ๊ฐ’์„ ๊ฐ–๊ฒŒ ๋œ๋‹ค. ์ด๋•Œ ๊ฐ’์ด ์Œ์ˆ˜๋ฅผ ๊ฐ€์ง€๋ฉด, Aij < didj/2m ์œผ๋กœ ์—ฐ๊ฒฐ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ํ•œ ์—ฃ์ง€ ๊ฐฏ์ˆ˜๋ณด๋‹ค ์‹ค์ œ ์—ฐ๊ฒฐ๋œ ๊ฒฝ์šฐ๊ฐ€ ๋” ์ž‘์€ ๊ฒƒ์œผ๋กœ ๋…ธ๋“œ u์™€ v ๊ฐ€ ์†ํ•œ cluster ๋Š” week group ์ด๋ผ ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๋ฐ˜๋ฉด 1์— ๊ฐ€๊นŒ์šด ์–‘์ˆ˜๊ฐ’์„ ๊ฐ€์ง„๋‹ค๋ฉด, ํ‰๊ท ์ ์ธ ์—ฐ๊ฒฐ๋ณด๋‹ค ๋” ๋งŽ์€ ์—ฃ์ง€์˜ ์—ฐ๊ฒฐ์ด ์ด๋ฃจ์–ด์ง„ ๊ฒƒ์ด๊ธฐ์— ํ•ด๋‹น cluster ๋Š” strong group ์ด๋ผ ํ•ด์„ํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค. 

 

 Modularity metric ์€ graph clustering metric ์œผ๋กœ ๊ฐ€์žฅ ์ž์ฃผ ์œ ์šฉํ•˜๊ฒŒ ์‚ฌ์šฉ๋˜๋Š” ์ง€ํ‘œ์— ํ•ด๋‹นํ•œ๋‹ค. 

 

 

 

โ‘ข  Spectral Modularity Maximization 

 

Modularity ๋ฅผ ์ตœ๋Œ€ํ™” ํ•˜๋Š” ๊ฒƒ์€, ๊ทธ๋ž˜ํ”„ ๋‚ด์— ์—ฃ์ง€๋ฅผ ์ „๋ถ€ ๋น„๊ตํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋‹คํ•ญ์‹œ๊ฐ„ ๋‚ด์— ํ’€ ์ˆ˜ ์—†๋Š” NP-hard ๋ฌธ์ œ๋ฅผ ๊ฐ€์ง€๊ฒŒ ๋œ๋‹ค. ๋”ฐ๋ผ์„œ ํšจ์œจ์ ์œผ๋กœ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด eigenvector ๋ฅผ ํ†ตํ•ด ๋ถ„ํ• ํ•˜๋Š” ๋ฐฉ๋ฒ•์ธ Spectral modularity ๋ฐฉ๋ฒ•์„ ๊ณ ์•ˆํ•˜๊ฒŒ ๋œ๋‹ค. 

 

 

โ–ธ C : cluster assignment matrix (nxk) , k is # of clusters

โ–ธ d : degree vector 

โ–ธ B : modularity matrix, B = A - dd' / 2m , A is adjacency matrix 

 

Modularity ๋ฅผ ์ตœ๋Œ€ํ™” ํ•˜๋ฉด์„œ Optimal ํ•œ C ๋Š” modularity matrix B ์˜ Top-k eigenvector ์ด๋‹ค. 

 

 

โ€ป eigenvalue, eigenvector 

ํšจ์œจ์ ์ธ ์—ฐ์‚ฐ์„ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋„๊ตฌ

 

 

 

 

 

 

5๏ธโƒฃ  Method 


 

โ‘   DMoN : Deep Modularity Networks 

 

• attributed graph clustering with GNN 

โ†ช Propse fully differentiable unsupervised clustering objective which optimize soft cluster assignments 

โ†ช Propse collapse regularization method 

 

• DMoN architecture ์˜ ํ•ต์‹ฌ์š”์†Œ 2๊ฐ€์ง€ 

 

(1) cluster assignments C ๋ฅผ encode ํ•˜๋Š” ๊ฒƒ 

 

 

์†Œํ”„ํŠธ๋งฅ์Šค ํ•จ์ˆ˜๋ฅผ ํ†ต๊ณผ์‹œ์ผœ ์ธ์ฝ”๋”ฉ ํ•จ์œผ๋กœ์จ ๋ฏธ๋ถ„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ๋งŒ๋“ ๋‹ค. soft ํ•œ ๋ฐฉ์‹์˜ ํด๋Ÿฌ์Šคํ„ฐ ํ• ๋‹น์„ ์œ„ํ•ด Graph Convolutional Network ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. A' ๋Š” normalized adjacency matrix (์ •๊ทœํ™”๋œ ์ธ์ ‘ํ–‰๋ ฌ) ๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ X ๋Š” node attributes (node features) ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. (n x s). D ๋Š” A ์˜ degree matrix ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. 

 

 

(2) C ๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š” objective function 

 

 

→ ( spectral modularity maximization + collapse reularization ) ์ด๋•Œ ์ •์น™ํ™” ๋ถ€๋ถ„์„ ์ถ”๊ฐ€ํ•œ ์ด์œ ๋Š” ์ตœ์ ํ™” ๋•Œ๋ฌธ!

 

โ–ธ B : modularity matrix (nxn), C : cluster assignment matrix (nxk), A : adjacency matrix (nxn)

 

 

โ–ธ Frobenius norm (Uclidean norm) ๋ฐฉ์‹์œผ๋กœ L2 norm ์ •์น™ํ™”์™€ ๋น„์Šทํ•œ ํ˜•ํƒœ๋ฅผ ๋ˆ๋‹ค. 

 

โ–ธ Tr( ) ์ˆ˜์‹์€ ํ–‰๋ ฌ A ์˜ ๋Œ€๊ฐ ์š”์†Œ๋“ค์„ ๋ชจ๋‘ ๋”ํ•˜๋Š” ์—ฐ์‚ฐ์„ ์˜๋ฏธํ•œ๋‹ค. 

 

 

 

โ‘ก  Collapse Regularization 

 

• spectral clustering ์˜ ๋‹จ์ ์„ ๊ทน๋ณตํ•ด์„œ ๋‚˜์˜จ ๊ทธ๋ž˜ํ”„ ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ธฐ๋ฒ•์ด MinCutPool 

 

MinCutPool ์˜ objective function ์€ ๋‘ ๊ฐ€์ง€ ์†์‹คํ•จ์ˆ˜๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋Š”๋ฐ, ์ฒซ๋ฒˆ์งธ๋Š” Cut loss Lc ์ด๊ณ  ๋‘๋ฒˆ์งธ๋Š” Orthogonality loss Lo ์ด๋‹ค. Lc ์˜ ์ˆ˜์‹์€ 'cluster ๋‚ด๋ถ€์˜ ์—ฐ๊ฒฐ์„ฑ' / 'graph ์ „์ฒด์˜ ์—ฐ๊ฒฐ์„ฑ' ์œผ๋กœ ๊ทธ๋ž˜ํ”„ ์ „์ฒด์˜ ์—ฐ๊ฒฐ์„ฑ๊ณผ ๋น„๊ตํ•ด ํ•ด๋‹น ํด๋Ÿฌ์Šคํ„ฐ ๋‚ด๋ถ€๊ฐ€ ์–ผ๋งˆ๋‚˜ ๊ฐ•ํ•˜๊ฒŒ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ๋Š”์ง€ ์•Œ๋ ค์ค€๋‹ค. ๊ฐ•ํ•˜๊ฒŒ ์—ฐ๊ฒฐ๋˜์–ด์žˆ๋Š” ๋…ธ๋“œ๋“ค์ด ์„œ๋กœ ํ•จ๊ป˜ cluster ๊ฐ€ ๋˜๋„๋ก ํ•œ๋‹ค. Cut loss ๋งŒ ๊ณ ๋ คํ•˜๊ฒŒ ๋œ๋‹ค๋ฉด node feature ๋ฅผ uniform ํ•˜๊ฒŒ ๋งŒ๋“ค์–ด local minima ์— ๋น ์งˆ ์œ„ํ—˜์ด ๋†’์•„์ง„๋‹ค. ๋”ฐ๋ผ์„œ Lo ๋ฅผ ๋„์ž…ํ•˜๋Š”๋ฐ, ์ด๋ฅผํ†ตํ•ด ํด๋Ÿฌ์Šคํ„ฐ ํ• ๋‹น์ด orthogonal ํ•˜๋„๋ก ํ•˜์—ฌ ๋ชจ๋“  ๋…ธ๋“œ๊ฐ€ ๋ชจ๋“  ํด๋Ÿฌ์Šคํ„ฐ์— ๊ฐ™์€ ์ •๋„๋กœ ์†ํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ , ํด๋Ÿฌ์Šคํ„ฐ์˜ ํฌ๊ธฐ๊ฐ€ ๋น„์Šทํ•˜๋„๋ก ์œ ๋„ํ•œ๋‹ค. 

 

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” MinCutPool ์˜ soft-orthogonality regularization ์ด softmax class assignment ์™€ ๊ฒฐํ•ฉ๋˜์—ˆ์„ ๋•Œ, C ๊ฐ’์˜ ๋ฒ”์ฃผ๋ฅผ ์™„์ „ํžˆ ์ œํ•œํ•˜์—ฌ loss ๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๋ฌธ์ œ๋ฅผ ์ œ๊ธฐํ•œ๋‹ค.

 

CORA dataset (๋…ผ๋ฌธ DB) ์— ๋Œ€ํ•ด 200 epoch ์˜ training ์„ ์ง„ํ–‰ํ•˜์˜€์„ ๋•Œ, Orthogonality regularization term ์ด MinCutPool ์˜ ์ตœ์ ํ™”๋ฅผ ์ขŒ์šฐํ•˜์—ฌ objective function ์ด ์ œ๋Œ€๋กœ ๋™์ž‘ํ•˜์ง€ ๋ชปํ•˜๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค. 

 

 

 

 

ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ณผ์ •์—์„œ MinCutPool objective function (joint optimization) ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ์™€ ๋‹จ์ˆœํžˆ feature orthogonalization ์„ ํ–ˆ์„ ๋•Œ๋ฅผ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ์ด๋‹ค. 5๊ฐœ์˜ dataset ์—์„œ 3๊ฐœ์˜ ๊ฒฝ์šฐ๊ฐ€ orthogonality ๋งŒ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ label correlation ์ด ๋” ์ข‹์•˜๋‹ค. ์ด์ฒ˜๋Ÿผ Regularization term ์€ ์ž์นซํ•˜๋‹ค ์„ฑ๋Šฅ์„ ์ขŒ์šฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค. 

 

๋”ฐ๋ผ์„œ DMoN ์€ MinCutPool ์˜ Lc + Lo joint optimization ๋ฐฉ์‹์ด ์•„๋‹Œ, collapse regularization ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•ด, Regularization ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•œ๋‹ค. ์ด๋ฅผํ†ตํ•ด ๋ชฉ์ ํ•จ์ˆ˜์˜ ์ตœ์ ํ™”๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ง„ํ–‰์‹œํ‚ฌ ์ˆ˜ ์žˆ๋„๋ก ํ•œ๋‹ค. 

 

 

โ‘ข  Theoretical insights 

 

• ์ œ์•ˆํ•œ objective function ์˜ ํƒ€๋‹น์„ฑ์„ ์ž…์ฆํ•˜๊ธฐ ์œ„ํ•ด, ํŠนํžˆ ์ƒˆ๋กญ๊ฒŒ ์ œ์‹œํ•˜๋Š” 'collapse regularization' ๋ฐฉ์‹์— ๋Œ€ํ•œ ์„ค๋ช…์„ ์œ„ํ•ด ์ •๋ฆฌํ•œ ์ด๋ก ๊ณผ ์ฆ๋ช…์„ ์ œ์‹œํ•œ๋‹ค. 

 

 

(1) ์ฒซ๋ฒˆ์งธ ์ด๋ก  

 

 

โœจ cluster assignments matrix C ๊ฐ€ positive modularity ๋ฅผ ๊ฐ€์ง„๋‹ค๋ฉด, objective score L(DMoN) ์€ ๊ฐ’์ด ๋” ์ž‘์•„์งˆ ๊ฒƒ์ด๋‹ค. (smaller loss) 

 

 

(2) ๋‘๋ฒˆ์งธ ์ด๋ก  (๊ทธ๋ž˜ํ”„ ํด๋Ÿฌ์Šคํ„ฐ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ผ๊ด€์„ฑ consistency ์— ๊ด€ํ•œ ๋‚ด์šฉ์„ ์†Œ๊ฐœ) 

 

 

โ†ช λn : average degree of the network (n ํ‘œ๊ธฐ๋Š” ํŠน์ • ๋…ธ๋“œ๋ฅผ ์˜๋ฏธํ•จ)

 

clustering ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด consistent ํ•˜๋‹ค๋Š” ๊ฒƒ์€ error rate ์„ vanishing ํ•˜๊ฒŒ ๋งŒ๋“ ๋‹ค๋Š” ๋œป์ด๋ฉฐ, compatiable ํ˜ธํ™˜ ๊ฐ€๋Šฅํ•œ ๋ชฉ์ ํ•จ์ˆ˜๋Š” ์ด๋Ÿฌํ•œ ์ผ๊ด€์„ฑ์˜ ํŠน์ง•์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค. DMoN ์€ collapse constraints part ๋ฅผ ๋„์ž…ํ•˜๋”๋ผ๋„ ์ผ๊ด€์ ์ธ prediction ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค. 

 

 

 

6๏ธโƒฃ  Experiments 

 


 

๐Ÿ‘€ ์ฝ”๋“œ : https://github.com/google-research/google-research/tree/master/graph_embedding/dmon

 

GitHub - google-research/google-research: Google Research

Google Research. Contribute to google-research/google-research development by creating an account on GitHub.

github.com

 

synthetic data (์ธ์œ„์ ์œผ๋กœ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ) ์™€ real-world data ๋ฐ์ดํ„ฐ์…‹์„ ํ™œ์šฉํ•ด ์‹คํ—˜์„ ์ง„ํ–‰ํ•˜์˜€์œผ๋ฉฐ, ํ‰๊ฐ€์ง€ํ‘œ๋กœ๋Š” graph clustering ์— ๊ด€ํ•œ ์ง€ํ‘œ์™€ label alignment ์— ๊ด€๋ จํ•œ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. 

 

 

 

๐Ÿ”น Datasets 

 

 

โ–ข Synthetic data 

 

์ธ์œ„์ ์œผ๋กœ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ

 

โ–ข Real datasets 

 

• |X| : node feature , |Y| : labels 

 

• Cora, Citeseer, Pubmed → Citation network 

 

๋…ธ๋“œ ์—ฃ์ง€ features labels
papars citation  bag of words of abstracts papaer topics

 

• Amazon PC, Photo → subsets of Amazon co-purchase graph for the pc and photo sections of the web 

 

๋…ธ๋“œ ์—ฃ์ง€ features labels
goods frequently purchased together bag of words of reviews product catecory 

 

 

• Coauthor CS, Coauthor Phys, Coauthor Med, Coauthor Chem, and Coauthor Eng  → co-authorship networks

 

๋…ธ๋“œ ์—ฃ์ง€ features labels
authors co-authored  collection of paper keywords most common fields of study

 

 

 

๐Ÿ”น Baselines 

 

๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ ํ˜น์€ ๋…ธ๋“œ ์†์„ฑ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ด์šฉํ•˜์—ฌ ํด๋Ÿฌ์Šคํ„ฐ๋ง์„ ํ•  ์ˆ˜ ์žˆ๋Š” baseline model ์„ ์‚ฌ์šฉํ•จ 

 

 

โ€ป SBM : ํŠน์ • ๋…ธ๋“œ๋ผ๋ฆฌ community structure ์„ ๊ฐ€์ง€๋Š” ๊ทธ๋ž˜ํ”„๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” generative model 

 

๊ฐ ๋ฒ ์ด์Šค๋ผ์ธ ๋ชจ๋ธ์— ๋Œ€ํ•˜์—ฌ, ๊ธฐ๋ณธ์ ์ธ GNN ๋ชจ๋ธ๊ตฌ์กฐ๋Š” one hidden layer with 512 neurons for real world datasets and 64 neurons for synthetic datasets ๋กœ ๊ณ ์ •ํ•˜๊ณ  ์ตœ๋Œ€ 16๊ฐœ์˜ cluster ๋กœ ํ• ๋‹นํ•  ์ˆ˜ ์žˆ๋„๋ก setting ์„ ์ง„ํ–‰ํ•œ ๋‹ค์Œ, Parameter tuning (e.g # of random walks, window size, loss coefficient, dropout...) ์„ ์ง„ํ–‰ํ–ˆ๋‹ค. 

 

 

 

๐Ÿ”น Metrics 

 

• graph based clustering metrics : graph modularity (๊ฐ’์ด ํด์ˆ˜๋ก ์ข‹์Œ), average cluster conductance (๊ฐ’์ด ์ž‘์„์ˆ˜๋ก ์ž˜ ํ˜•์„ฑ๋œ ํด๋Ÿฌ์Šคํ„ฐ๋ผ ๋ณผ ์ˆ˜ ์žˆ์Œ)

label based metrics : normalized mutual information (๊ฐ’์ด ํด์ˆ˜๋ก ์ข‹์Œ), pairwise F1 score 

 

 

โ•  cf) ๋ชจ๋“  ์ง€ํ‘œ๊ฐ’์— 100์„ ๊ณฑํ•˜์—ฌ ์„ฑ๋Šฅ ๋น„๊ต๋ฅผ ์ง„ํ–‰ํ•จ 

 

 

๐Ÿ”น Results 

 

(1) Synthetic data 

 

DMoN ์˜ ์„ฑ๋Šฅ์ด ๋‹ค๋ฅธ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋ณด๋‹ค ์šฐ์„ธํ•จ

 

 

(2) Real world data 

 

 

 

๐Ÿ‘‰ ๊ตต์€ ํ™œ์ž๊ฐ€ ๊ฐ€์žฅ best score ๋ฅผ ๋ณด์ธ ๋ถ€๋ถ„์ด๋ผ ๋ณผ ์ˆ˜ ์žˆ์Œ 

 

 

 

 

7๏ธโƒฃ  Conclusions 


 

โ‘  DMoN ์€ attributed graph clustering ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ, unsupervised objective ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ด ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•œ๋‹ค. 

 

โ‘ก DMoN ์€ graph ๊ตฌ์กฐ์™€ node feature ๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•˜์—ฌ ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, synthetic data ์™€ real world data ์—์„œ ๋ชจ๋‘ baseline model ๋ณด๋‹ค ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. 

 

 

 

728x90

๋Œ“๊ธ€