Graph attention network 기반 멀티뷰 그래프 임베딩 모델 연구

국문 초록 (Abstract)

소셜 네트워크, 인용 네트워크 등 현실 세계는 주체 간의 다양한 연결관계로 구성되어 있다. 그래프는 이러한 연결관계를 반영하는 데이터 형태이며, 이러한 데이터를 기계 처리가 가능하게 만들기 위해서는 그래프 임베딩 과정이 필요하다. 본 연구는 네트워크 임베딩 방법론을 사용하여 그래프 내의 다양한 관계를 반영하여, 더 높은 표현력을 가지는 노드를 학습함으로써 노드 분류와 그래프 분류, 그래프 클러스터링에 있어서의 성능을 높이고자 한다. 노드 분류에 쓰인 데이터로는 Amazon 상품구매 데이터와 IMDB 프로그램 데이터를 사용하였으며, 그래프 분류와 그래프 클러스터링에 있어서는 통신사 3사 고객들의 IPTV 시청기록 데이터를 기반으로 하였다. 이러한 태스크들에 있어 적용한 방법론은 Deep Multiplex Graph Infomax(DMGI) 알고리즘과 Multi-view Graph Attention Network(MGAT) 알고리즘을 기반으로 한다. DMGI 방법론은 노드에 대한 라벨이 없이도 임베딩이 가능한 비지도 학습 방법이지만, 그래프 관계 간의 중요도를 학습하지 못한다는 단점이 있다. 이에 따라 Multi-view Graph Attention Network(MGAT) 방법론을 사용하여 노드 임베딩 시 관계 간의 중요도를 반영하였다. 그 결과 임베딩 벡터의 표현력을 높여 학습된 모델을 활용하여, 노드 분류와 그래프 클러스터링에 대한 성능을 높일 수 있었다.

다국어 초록 (Multilingual Abstract)

The real world, such as social networks and citation networks, is
composed of various connections between subjects. A graph is a form of
data that reflects these connections, and a graph embedding process is
required to make such data machine-processable. This study aims to improve
performance in node classification, graph classification, and graph
clustering by learning nodes with higher expressive power by reflecting
various relationships in the graph using the network embedding methodology.
Amazon product purchase data and IMDB program data were used for node
classification, and IPTV viewing record data of customers of three
telecommunication companies was used for graph classification and graph
clustering.
The methodology applied to these tasks is based on the Deep Multiplex
Graph Infomax (DMGI) algorithm and the Multi-view Graph Attention Network
(MGAT) algorithm. The DMGI methodology is an unsupervised learning method
that enables embedding without labeling nodes, but has a disadvantage in
that it cannot learn the importance between relationships in multi-view
graph. Accordingly, the multi-view graph attention network (MGAT)
methodology was used to reflect the importance of relationships when
embedding nodes. As a result, it was possible to improve the performance
of node classification and graph clustering by using the learned model by
increasing the expressiveness of the embedding vector.

목차 (Table of Contents)

• < 그림 차례 > ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․iii
• < 표 차례 > ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ iv
• < 국문 요약 > ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․v
• 1. 서론 ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․1
• 2. 관련 연구 ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․5
• 2.1 Attention-based network embedding․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ 5
• 2.2 Graph neural networks(GNNs) ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․6
• 2.3 Multi-view network embedding․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ 7
• 2.4 기존 연구의 한계점․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ 9
• 3. 연구 방안․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․10
• 3.1 Deep Graph Infomax․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․11
• 3.2 Deep Multiplex Graph Infomax․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․12
• 3.3 Multi-view graph attention network ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․13
• 3.4 Node-level task에서의 제안 모델․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․14
• 3.4.1 Relation specific node embedding․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․16
• 3.4.2 MGAT based consensus node embedding․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․18
• 3.5 Graph-level task에서의 제안모델․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․21
• 3.5.1 Relation specific node embedding․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․21
• 3.5.2 MGAT based consensus node embedding & pooling․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․22
• 4. 실험 및 결과 ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․24
• 4.1 실험 데이터 및 전처리․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․24
• 4.2 Node-level task․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․27
• 4.2.1 Node classification․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․ 27
• 4.3 Graph-level task․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․29
• 4.3.1 Graph clustering․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․29
• 4.3.2 Graph classification․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․32
• 5. 결론 ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․33
• < 참고 문헌 > ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․34
• < 영문 요약 > ․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․․36