Exploit Label Embeddings for Enhancing Network Classification

Learning representations for network has aroused great research interests in recent years. Existing approaches embed vertices into a low dimensional continuous space which encodes local or global network structures. While these methods show improvements over traditional representations, they do not utilize the label information until the learned embeddings are used for training classifier. That is, the process of representation learning is separated from the labels and thus is unsupervised. In this paper, we propose a novel method which learns the embeddings for vertices under the supervision of labels. The key idea is to regard the label as the context of a vertex. More specifically, we attach a true or virtual label node for each training or test sample, and update the embeddings for vertices and labels to maximize the probability of both the neighbors and their labels in the context. We conduct extensive experiments on three real datasets. Results demonstrate that our method outperforms the state-of-the-art approaches by a large margin.