Abstract
Learning representations for dynamic graphs is fundamental as it supports numerous graph analytic tasks such as dynamic link prediction, node classification, and visualization. Real-world dynamic graphs are continuously evolved where new nodes and edges are introduced or removed during graph evolution. Most existing dynamic graph representation learning methods focus on modeling dynamic graphs with fixed nodes due to the complexity of modeling dynamic graphs, and therefore, cannot efficiently learn the evolutionary patterns of real-world evolving graphs. Moreover, existing methods generally model the structural information of evolving graphs separately from temporal information. This leads to the loss of important structural and temporal information that could cause the degradation of predictive performance of the model. By employing an innovative neural network architecture based on graph attention networks and temporal convolutions, our framework jointly learns graph representations contemplating evolving graph structure and temporal patterns. We propose a deep attention model to learn low-dimensional feature representations which preserves the graph structure and features among series of graph snapshots over time. Experimental results on multiple real-world dynamic graph datasets show that, our proposed method is competitive against various state-of-the-art methods.