Article Source

• TGN: Temporal Graph Networks for Dynamic Graphs
• Blog post

TGN: Temporal Graph Networks for Dynamic Graphs

Video for MLSS 2020 Tübingen presenting TGN: Temporal Graph Networks for Dynamic Graphs.

TGN: Key Insights

• Memory: to store the dynamic state of each node. It’s updated each time a node is involved in an interaction.
• Graph Embedding: when computing the temporal embedding for a node, aggregate the memories of the neighbors.

TGN: Modules and Training

Computations performed by TGN on a batch of time-stamped interactions.

Flow of operations of TGN used to train the memory-related modules.

Conclusion

• Dynamic graphs are very common, but have received little attention so far.
• We propose TGN, which generalizes existing models and achieves SOTA results on a variety of benchmarks.
• The ablation study shows the importance of the different modules.