A scalable algorithm for graph-based active learning

In many learning tasks, to obtain labeled instances is hard due to heavy cost while unlabeled instances can be easily collected. Active learners can significantly reduce labeling cost by only selecting the most informative instances for labeling. Graph-based learning methods are popular in machine learning in recent years because of clear mathematical framework and strong performance with suitable models. However, they suffer heavy computation when the whole graph is in huge size. In this paper, we propose a scalable algorithm for graph-based active learning. The proposed method can be described as follows. In the beginning, a backbone graph is constructed instead of the whole graph. Then the instances in the backbone graph are chosen for labeling. Finally, the instances with the maximum expected information gain are sampled repeatedly based on the graph regularization model. The experiments show that the proposed method obtains smaller data utilization and average deficiency than other popular active learners on selected datasets from semi-supervised learning benchmarks.