Deep Multi-Attributed-View Graph Representation Learning

Graph representation learning aims at mapping a graph into a lower-dimensional feature space. Deep attributed graph representation, utilizing deep learning models on the graph structure and attributes, shows its significance in mining complex relational data. Most existing deep attributed graph representation models assume graph attributes in a single-attributed view. However, rich information in real-world applications demands the ability to handle multiple attributed views. For example, in social network users’ profiles and posts represent two distinct attributed views. A single-attributed view or a simple ensemble of them fails to represent the rich information and complex relations therein. To confront this challenge, this paper proposes a novel deep unsupervised graph representation learning model, called Multi-attributed-view graph Convolutional AutoEncoder (MagCAE). MagCAE learns the node-pairwise proximity among multi-attributed views and node embeddings, across which a novel loss function is designed to preserve the node-pairwise likelihood. An aggregation layer is specially developed in MagCAE to optimize the weights of embeddings on multi-attributed views. The extensive experiments on four datasets demonstrate the superiority of MagCAE over twelve baselines.