TY - GEN
T1 - C2N-ABDP: Cluster-to-Node Attention-Based Differentiable Pooling
AU - Ye, Rongji
AU - Cui, Lixin
AU - Rossi, Luca
AU - Wang, Yue
AU - Xu, Zhuo
AU - Bai, Lu
AU - Hancock, Edwin R.
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2023/9
Y1 - 2023/9
N2 - Graph neural networks have achieved state-of-the-art performance in various graph based tasks, including classification and regression at both node and graph level. In the context of graph classification, graph pooling plays an important role in reducing the number of graph nodes and allowing the graph neural network to learn a hierarchical representation of the input graph. However, most graph pooling methods fail to effectively preserve graph structure information and node feature information when reducing the number of nodes. At the same time, the existing hierarchical differentiable graph pooling methods cannot effectively calculate the importance of nodes and thus cannot effectively aggregate node information. In this paper, we propose an attention-based differentiable pooling method, which aggregates nodes into clusters when reducing the scale of the graph, uses singular value decomposition to calculate cluster information during the aggregation process, and captures node importance information through a novel attention mechanism. The experimental results show that our approach outperforms competitive models on benchmark datasets.
AB - Graph neural networks have achieved state-of-the-art performance in various graph based tasks, including classification and regression at both node and graph level. In the context of graph classification, graph pooling plays an important role in reducing the number of graph nodes and allowing the graph neural network to learn a hierarchical representation of the input graph. However, most graph pooling methods fail to effectively preserve graph structure information and node feature information when reducing the number of nodes. At the same time, the existing hierarchical differentiable graph pooling methods cannot effectively calculate the importance of nodes and thus cannot effectively aggregate node information. In this paper, we propose an attention-based differentiable pooling method, which aggregates nodes into clusters when reducing the scale of the graph, uses singular value decomposition to calculate cluster information during the aggregation process, and captures node importance information through a novel attention mechanism. The experimental results show that our approach outperforms competitive models on benchmark datasets.
KW - Attention
KW - Graph neural network
KW - Graph pooling
UR - http://www.scopus.com/inward/record.url?scp=85171597740&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-42795-4_7
DO - 10.1007/978-3-031-42795-4_7
M3 - Conference article published in proceeding or book
AN - SCOPUS:85171597740
SN - 9783031427947
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 70
EP - 80
BT - Graph-Based Representations in Pattern Recognition - 13th IAPR-TC-15 International Workshop, GbRPR 2023, Proceedings
A2 - Vento, Mario
A2 - Foggia, Pasquale
A2 - Carletti, Vincenzo
A2 - Conte, Donatello
PB - Springer Science and Business Media Deutschland GmbH
T2 - 13th IAPR-TC-15 International Workshop on Graph-Based Representations in Pattern Recognition, GbRPR 2023
Y2 - 6 September 2023 through 8 September 2023
ER -