Behavior-Based Collective Classification in Sparsely Labeled Networks

Classification in sparsely labeled networks is challenging to traditional neighborhood-based methods due to the lack of labeled neighbors. In this paper, we propose a novel behavior-based collective classification (BCC) method to improve the classification performance in sparsely labeled networks. In BCC, nodes' behavior features are extracted and used to build latent relationships between labeled nodes and unknown ones. Since mining the latent links does not rely on the direct connection of nodes, decrease of labeled neighbors will have minor effect on classification results. In addition, the BCC method can also be applied to the analysis of networks with heterophily as the homophily assumption is no longer required. Experiments on various public data sets reveal that the proposed method can obtain competing performance in comparison with the other state-of-the-art methods either when the network is labeled sparsely or when homophily is low in the network.