Exploring a link between network topology and active learning

In many online networking platforms whether they are online social networks or academic citation networks, community label memberships are applied to the nodes to generalize overlaps based upon common features or associations. The relatively recent methods in graph convolutional neural networks has provided new tools to infer community labels of nodes but they still depend upon a labeled dataset to be provided. Obtaining these labels can be costly and methods to reduce the required number of labels needed can speed up the process and reduce costs. This study explores different strategies for selecting nodes to be used as training data showing which strategies work better or worse and on different percentages of the network's nodes. An unsupervised approach of deciding the best active learning sampling direction (i.e. ascending or descending selection of nodes in terms of importance) procedure is derived by fundamental network properties. The conclusion is supported on both simulated and real data.