Label-expanded manifold regularization for semi-supervised classification

Manifold regularization (MR) provides a powerful framework for semi-supervised classification, which propagates labels from the labeled instances to unlabeled ones so that similar instances over the manifold have similar classification outputs. However, labeled instances are randomly located. Label propagation from those instances to their neighbors may mislead the classification of MR. To address this issue, in this paper we develop a novel label-expanded MR framework (LE_MR for short) for semi-supervised classification. In LE_MR, a clustering strategy such as KFCM is first adopted to discover the high-confidence instances, i.e., instances in the central region of clusters. Then those instances along with the cluster indices are adopted to expand the labeled instances set. Experiments show that LE_MR obtains encouraging results compared to state-of-the-art semi-supervised classification methods.