To Actively Initialize Active Learning

Yazhou Yang, Marco Loog

Abstract: Highlights•The initialization problem of active learning is addressed, which is how to find a set of labeled samples which contains at least one instance per category.•A new active initialization criterion, the Nearest Neighbor Criterion, is proposed for the initialization task.•The impacts of different initialization strategies on the whole active learning process are further investigated.•Excellent performance of the proposed method in comparison with state-of-the-art approaches is demonstrated.