Pre-training on dynamic graph neural networks

The pre-training of the graph neural network model is to learn general characteristics large-scale graphs or similar type usually through a self-supervised method, which allows work even when node labels are missing. However, existing methods do not take temporal information edge generation and evolution process into consideration. To address this issue, paper proposes method based on dynamic networks (PT-DGNN), uses task simultaneously structure, semantics, features graph. mainly includes two steps: 1) subgraph sampling, 2) using tasks. former preserves local time-aware structure original by sampling latest frequently interacting nodes. latter observed edges predict unobserved capture evolutionary network. Comparative experiments three realistic datasets show that proposed achieves best results link prediction fine-tuning ablation study further verifies effectiveness above steps.