Robust Classification of Information Networks by Consistent Graph Learning.

Graph regularization-based methods have achieved great success for network classification by making the label-link consistency assumption, i.e., if two nodes are linked together, they are likely to belong to the same class. However, in a real-world network, there exist links that connect nodes of different classes. These inconsistent links raise a big challenge for graph regularization and deteriorate the classification performance significantly. To address this problem, we propose a novel algorithm, namely Consistent Graph Learning, which is robust to the inconsistent links of a network. In particular, given a network and a small number of labeled nodes, we aim at learning a consistent network with more consistent and fewer inconsistent links than the original network. Since the link information of a network is naturally represented by a set of relation matrices, the learning of a consistent network is reduced to learning consistent relation matrices under some constraints. More specifically, we achieve it by joint graph regularization on the nuclear norm minimization of consistent relation matrices together with ℓ1-norm minimization on the difference matrices between the original relation matrices and the learned consistent ones subject to certain constraints. Experiments on both homogeneous and heterogeneous network datasets show that the proposed method outperforms the state-of-the-art methods.