Abstract

We present very efficient active learning algorithms for link classification in signed networks. Our algorithms are motivated by a stochastic model in which edge labels are obtained through perturbations of a initial sign assignment consistent with a two-clustering of the nodes. We provide a theoretical analysis within this model, showing that we can achieve an optimal (to whithin a constant factor) number of mistakes on any graph G = (V, E) such that |E| = by querying edge labels. More generally, we show an algorithm that achieves optimality to within a factor of O(k) by querying at most order of edge labels. The running time of this algorithm is at most of order