Representation Learning on Graphs by Integrating Content and Structure Information

The problem of representation learning on graph can be difficult due to limited knowledge of training data and large presence of missing edges. Real-world social networks do not provide complete information about the network due to hidden information and privacy constraints. In such scenarios, typical representation learning methods are not able to capture network information effectively. In order to make them more useful, any available feature information can be used in addition to the network structure. In this paper, we aim to learn better representations by exploiting both content (or feature) information of nodes and structural information of the network. Our approach leverages generative adversarial networks to learn embedding for generator and discriminator in a minimax game. While the generator estimates the neighborhood of a node, the discriminator distinguishes between the presence or absence of a link for a pair of nodes. We demonstrate the effectiveness of our approach on five real-world publicly available datasets on the problems of link prediction and node classification. On both tasks, we achieve significant gains, outperforming current state-of-the-art methods by considerable margins. Our code is available on Github ¹ .