DynGraphSAGE: An Inductive Embedding Method for Dynamic Graphs

Learning low-dimensional representations of nodes on large graphs has recently become increasingly popular due to its wide applications ranging from biology and physics to social networks and recommendation systems. Despite the plethora of different models for representation learning on graphs, few approaches have been proposed for dealing with graphs that are dynamic in nature (e.g., evolving nodes, connectivity or features over time). In this paper, we present DynGraphSAGE, a general inductive framework extended from GraphSAGE that generates node representations on dynamic graphs in an efficient manner. Specifically, DynGraphSAGE computes node representations by employing the aggregation and message propagation scheme in both spatial and temporal neighbors. Besides the function that samples and aggregates features from a node's spatial neighborhood, we also introduce and learn an additional function that generates embeddings through aggregating from the node's temporal neighborhood. The proposed DynGraphSAGE algorithm has achieved state-of-the-art results across three established dynamic graph representation benchmarks: the Enron, UCI, and ML-10m network datasets.