Active learning via collective inference in network regression problems

Abstract

Active learning is a promising machine learning paradigm for querying oracles and obtaining actual labels for particular examples. Its goal is to decrease the number of labels needed, in order to learn a predictive model able to achieve a high level of accuracy. It may turn out to be advantageous in several regression problems where scarce labels can be acquired. A novel active learning algorithm for regression problems in network data is defined. This algorithm performs active learning by taking into account explicitly the correlation property of network data, which makes the labels of linked nodes related to each other. Specifically it resorts to collective inference, in order to accommodate the data correlation in the active selection of the network nodes labeled by oracles. The empirical study proves that the proposed combination of active learning and collective inference can actually boost regression performances in various network domains.