Active learning allows semi-supervised clustering algorithms to sol-licit domain experts to retrieve a few set of class labels (or seeds) to improve theirefficiency or the relevance of their results. However, some recent studies showthat, even in the case of a good answer from the domain expert, semi-supervisedclustering can see their performances drop with badly chosen seeds. Until now,only few works address the problem of determining the best queries for a cluster-ing algorithm in an active learning context, and most of these studies are limitedbecause of their hypothesis on the size and the shape of expected clusters. In thispaper, we propose a new active seed selection algorithm that makes no hypothe-sis on the underlying data distribution. Experiments conducted on real data setsshow the efficiency of this new approach compared to existing ones.

Tải file Active seeds selection with a k-nearest neighbors graph tại đây