Inductive Representation Learning via CNN for Partially-Unseen Attributed Networks

Network embedding aims to map a complex network into a low-dimensional vector space while maximally preserving the properties of the original network. An attributed network is a typical real-world network that models the relationships and attributes of real-world entities. Its analysis is of great significance in many applications. However, most such networks are incomplete with partially-known attributes, links and labels. Traditional network embedding methods are designed for a complete network and cannot be applied to a network with incomplete information. Thus, this work proposes an inductive embedding model to learn the robust representations for a partially-unseen attributed network. It is designed based on a multi-core convolutional neural network and a semi-supervised learning mechanism, which can preserve the properties of such a network and generate the effective representations for unseen nodes in a model training process. We evaluate its performance on the task of inductive node classification and community detection via three real-world attributed networks. Experimental results show that it significantly outperforms the state-of-the-art.