Graph Active Learning at Subgraph Granularity

Graph active learning algorithms can reduce the amount of labeling and improve the applicability of graph neural networks. However, existing graph active learning algorithms are mainly performed at the node granularity. Those setting does not hold to datasets that are sensitive to edge attributes. To solve this problem, we propose a graph active learning algorithm at subgraph granularity. The algorithm tackles two critical challenges: how to estimate the expected labeling value of subgraphs and how to search for high-value subgraphs in the whole graph efficiently. For the first challenge, we evaluate the expected labeling values of subgraphs based on heuristic metrics, including uncertainty, representativeness, centrality, and diversity. Among them, uncertainty cannot be measured directly. Therefore, we measure subgraph inner cohesion by GNN attention weights and estimate uncertainty based on it. For subgraph search, we propose an efficient subgraph search algorithm. The proposed algorithm includes a simulated annealing search algorithm for a single subgraph and beam search with subgraph-effective reception field algorithms for multiple subgraphs. Experiments demonstrate that the subgraph granularity active learning algorithm proposed in this paper can achieve great results on edge-sensitive datasets.