Aspect-Level Attributed Network Embedding via Variational Graph Neural Networks

Attributed information network embedding (AINE) has been widely used in network analysis. Existing AINE methods mainly focus on preserving network proximities and minimizing the reconstruction loss of node attribute information from a single aspect. However, complex network data may stem from different aspects. For example, a social network may consist of working relationship networks, alumni associations and so on. In this paper, we propose a novel model, called Aspect-level Attributed Network Embedding (AANE), to embed nodes by learning different aspect-level information. Specifically, we use a transform matrix to model aspect-level network topological structure and node attributes. Then, we leverage graph neural networks to learn aspect-level embedding. To learn a robust representation, we aggregate different aspect-level embeddings via the attention mechanism in a variational manner. Experimental results on four real-world network datasets demonstrate that AANE outperforms the state-of-the-art network embedding methods.