Active learning is concerned with the design of prediction algorithms which select labeled training samples that are expected to best reduce the label prediction error. An active learner queries those labels (from an annotator\oracle) either sequentially or in batch mode and re-learns a model after each query step. It has been shown that the active selection of training samples expedites the learning process over traditional semi-supervised learning (where arbitrary\random training sets are given), and is valuable in many realistic settings in which labels are scarce or expensive to obtain.
I will present two graph-based active learners for the task of classification and for the task of community-detection in networks, time permitting. The first learner is based on a label diffusion framework in graphs for classification in which diffusion kernels are adapted to the geometry of the labeling function to be reconstructed. This adaptation allows for a very efficient active-query criterion for most uncertain data points and introduces a 5000X speed-up over state-of-the-art active learners. The second active learner is based on a maximum-likelihood framework for the recovery of communities in networks where the nodes are assumed to have labels associated with them (permutation invariant). We investigate the performance of active community detection in light of the known (unsupervised) community recovery bounds for the benchmarked Block-Stochastic Model (SBM). We demonstrate how in practice active learning allows to achieve below-detection-bound recovery for SBM. I will present applications involving detection of organizational and social community structure based on in-building WiFi-based mobile-device collocation networks, and more.