DynGraphGAN: Dynamic Graph Embedding via Generative Adversarial Networks

Abstract

Graphs have become widely adopted as a means of representing relationships between entities in many applications. These graphs often evolve over time. Learning effective representations preserving graph topology, as well as latent patterns in temporal dynamics, has drawn increasing interests. In this paper, we investigate the problem of dynamic graph embedding that maps a time series of graphs to a low dimensional feature space. However, most existing works in the field of dynamic representation learning either consider temporal evolution of low-order proximity or treat high-order proximity and temporal dynamics separately. It is challenging to learn one single embedding that can preserve the high-order proximity with long-term temporal dependencies. We propose a Generative Adversarial Networks (GAN) based model, named DynGraphGAN, to learn robust feature representations. It consists of a generator and a discriminator trained in an adversarial process. The generator generates connections between nodes that are represented by a series of adjacency matrices. The discriminator integrates a graph convolutional network for high-order proximity and a convolutional neural network for temporal dependency to distinguish real samples from fake samples produced by the generator. With iterative boosting of the performance of the generator and discriminator, node embeddings are learned to present dynamic evolution over time. By jointly considering high-order proximity and temporal evolution, our model can preserve spatial structure with temporal dependency. DynGraphGAN is optimized on subgraphs produced by random walks to capture more complex structural and temporal patterns in the dynamic graphs. We also leverage sparsity and temporal smoothness properties to further improve the model efficiency. Our model demonstrates substantial gains over several baseline models in link prediction and reconstruction tasks on real-world datasets.