.png){kind=icon}
.png){kind=icon}
Return
Multi-view parallel graph pooling
.png){kind=icon}
PRE
AI
DOI：10.1007/s41060-023-00476-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph pooling is a crucial operation in graph neural networks (GNNs) for down-sampling. It is noted that existing methods for graph pooling suffer from two main issues. First, pooling methods based on node dropping only evaluate the importance of nodes from a single perspective, such as attention or node distance. However, the importance scores of a node are different when evaluating it from different aspects, and thus the importance of a node cannot be evaluated comprehensively only based on one single view. Second, necessary information about graph structure will be lost when nodes with lower scores are discarded. It can even result in the coarsened graph being too sparse and further lead to poor performance in related learning tasks, such as graph classification. To address these issues, we propose an attention-based multi-view parallel graph pooling method. Specifically, to comprehensively evaluate the importance of nodes, we propose to evaluate the importance node from its features, local topology structure, and global topology structure via the attention mechanism. Moreover, to alleviate the problem of information loss caused by node discarding in graph pooling, we introduce the concept of multi-view parallel pooling, which conducts graph pooling from node features, local topology, and global topology structure three views, respectively, and then integrates the three generated coarsened graphs and obtains an informative graph. Experimental results on four benchmark datasets demonstrate the effectiveness of our proposed method for graph classification.
Keywords:
Graph pooling
Graph classification
Graph neural network
Attention mechanism
Journal
I
IF:
2.8
Papers:
1.0K
Citations:
1.3K