Multi-View Learning of Network Embedding

Abstract

In recent years, network representation learning on complex information networks attracts more and more attention. Scholars usually use matrix factorization or deep learning methods to learn network representation automatically. However, existing methods only preserve single feature of networks. How to effectively integrate multiple features of network is a challenge. To tackle this challenge, we propose an unsupervised learning algorithm named Multi-View Learning of Network Embedding. The algorithm preserves multiple features that including vertex attribute, network global and local topology structure. Features are treated as network views. We use a variant of convolutional neural networks to learn features from these views. The algorithm maximizes the correlation between different views by canonical correlation analysis, and learns the embedding that preserve multiple features of networks. Comprehensive experiments are conducted on five real networks. We demonstrate that our method can better preserve multiple features and outperform baseline algorithms in community detection, network reconstruction and visualization.