Partition and Learned Clustering with joined-training: Active learning of GNNs on large-scale graph

• Published in: Knowledge-based systems 2022-12, Vol.258, p.110050, Article 110050
• Main Authors: Gao, Jian, Wu, Jianshe, Zhang, Xin, Li, Ying, Han, Chunlei, Guo, Chubing
• Format: Article
• Language: English
• Subjects: Active learning; Graph neural network; Graph partition; Node classification; Semi-supervised learning
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2022.110050

Graph neural networks (GNNs) have recently achieved impressive progress on graph-based semi-supervised learning. The active learning of GNNs aims to select a small number of representative nodes to train a high-performance GNN. However, existing active learning methods of GNNs have the following unsolved problems: (1) only consider nodes distribution in embedding space, without considering the distribution on graph structure; (2) difficult to be applied to large-scale graphs; (3) cluster-based methods do not consider topological structure during clustering; (4) the search space is inconsistent with the classification space. To address these problems, we propose a novel cluster-based active learning method of GNNs on large-scale graphs, called Partition and Learned Clustering with Joined-training (PLCJ). PLCJ first partitions the graph into several subgraphs, then clusters in each subgraph, and the cluster centers are selected. The partition makes the selected nodes evenly distributed on the structure and makes PLCJ can be applied to a large-scale graph. Moreover, PLCJ uses Learned Clustering, which clusters nodes considering both embedding and structural information. In the classification stage, PLCJ employs a joined-training skill to ensure the closeness of search and classification space. Experiments show that PLCJ can achieve better performance than state of the art active learning methods on graphs with only a few percent of the running time.