Semisupervised Classification With Cluster Regularization

Semisupervised classification (SSC) learns, from cheap unlabeled data and labeled data, to predict the labels of test instances. In order to make use of the information from unlabeled data, there should be an assumed relationship between the true class structure and the data distribution. One assumption is that data points clustered together are likely to have the same class label. In this paper, we propose a new algorithm, namely, cluster-based regularization (ClusterReg) for SSC, that takes the partition given by a clustering algorithm as a regularization term in the loss function of an SSC classifier. ClusterReg makes predictions according to the cluster structure together with limited labeled data. The experiments confirmed that ClusterReg has a good generalization ability for real-world problems. Its performance is excellent when data follows this cluster assumption. Even when these clusters have misleading overlaps, it still outperforms other state-of-the-art algorithms.