Towards Active Learning on Graphs: An Error Bound Minimization Approach

Active learning on graphs has received increasing interest in the past years. In this paper, we propose a \textit{nonadaptive} active learning approach on graphs, based on generalization error bound minimization. In particular, we present a data-dependent error bound for a graph-based learning method, namely learning with local and global consistency (LLGC). We show that the empirical transductive Rademacher complexity of the function class for LLGC provides a natural criterion for active learning. The resulting active learning approach is to select a subset of nodes on a graph such that the empirical transductive Rademacher complexity of LLGC is minimized. We propose a simple yet effective sequential optimization algorithm to solve it. Experiments on benchmark datasets show that the proposed method outperforms the state-of-the-art active learning methods on graphs.