A Survey on Dynamic Graph Neural Networks Modeling

Compared with the Static Graph Neural Networks(SGNNs), the Dynamic Graph Neural Networks(DGNNs) have an additional temporal dimension, which can evolve along the time dimension, so DGNNs have more powerful and flexible modeling capabilities. DGNNs are widely used in the classical graph domain, such as social network recommendations, chemical molecular structure predictions and so on. They are also succeeding in areas such as social event prediction, video classification, traffic prediction. Although the advantages of DGNNs have been verified in various aspects, there are still relatively few relevant review studies. This paper describes the reason why DGNNs can process temporal data, which lays a foundation for subsequent understanding of various DGNNs models. Based on temporal models and GNN models, the whole design idea of DGNNs is provided for readers through detailed description of encoding and decoding process and link predictions of various DGNNs. Finally, we are dedicated to mining the commonness of DGNN framework from different task scenarios, and presenting a detailed DGNNs overview for readers.