Semi-Supervised network embedding

Network embedding aims to learn a distributed representation vector for each node in a network, which is fundamental to support many data mining and machine learning tasks such as node classification, link prediction, and social recommendation. Current popular network embedding methods normally first transform the network into a set of node sequences, and then input them into an unsupervised feature learning model to generate a distributed representation vector for each node as the output. The first limitation of existing methods is that the node orders in node sequences are ignored. As a result some topological structure information encoded in the node orders cannot be effectively captured by such order-insensitive embedding methods. Second, given a particular machine learning task, some annotation data can be available. Existing network embedding methods are unsupervised and are not effective to incorporate the annotation data to learn better representation vectors. In this paper, we propose an order sensitive semi-supervised framework for network embedding. Specifically, we first propose an novel order sensitive network embedding method: StructuredNE to integrate node order information into the embedding process in an unsupervised manner. Then based on the annotation data, we further propose an semi-supervised framework SemNE to modify the representation vectors learned by StructuredNE to make them better fit the annotation data. We thoroughly evaluate our framework through three data mining tasks (multi-label classification, network reconstruction and link prediction) on three datasets. Experimental results show the effectiveness of the proposed framework.