Skip to content
/ tgn-aa Public
forked from twitter-research/tgn

TGN-AA: Temporal Graph Networks with attention-based aggregator

License

Notifications You must be signed in to change notification settings

LyoAI/tgn-aa

 
 

Repository files navigation

TGN: Temporal Graph Networks [arXiv, YouTube, Blog Post]

Dynamic Graph TGN

What's Difference between tgn and tgn-aa

Multiple choices can be considered for implementing the Message Aggregator module. While the original paper only considered two efficient non-learnable solutions: most recent message (keep only most recent message for a given node) and mean message (average all messages for a given node), our tgn-aa design a learnable attention-based aggregation to aggregate the messages from multiple events for nodes in the same batch

TGN-AA

What's changed

|-modules
|- |-message_aggregator.py add AttentionMessageAggregator
|-train_self_supervised.py add attention for argument --aggregator
|-train_supervised.py add attention for argument --aggregator

Preprocess the data

We use the dense npy format to save the features in binary format. If edge features or nodes features are absent, they will be replaced by a vector of zeros.

python utils/preprocess_data.py --data wikipedia --bipartite

Training on Wikipedia Dataset

### tgn-aa
# TGN-attn with attention aggregator: Self-Supervised learning on the wikipedia dataset
## unlearnable
python train_self_supervised.py --aggregator attention --use_memory --prefix tgn-attn --n_runs 10
## learnable
python train_self_supervised.py --aggregator attention --learnable --use_memory --prefix tgn-attn --n_runs 10 --add_cls_token


### Baselines
# Jodie
python train_self_supervised.py --use_memory --memory_updater rnn --embedding_module time --prefix jodie_rnn --n_runs 10

# Jodie
python train_supervised.py --use_memory --memory_updater rnn --embedding_module time --prefix jodie_rnn --n_runs 10

About

TGN-AA: Temporal Graph Networks with attention-based aggregator

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%