This example is used as a demo in the lightening talk about graphs at Bumble X Data science festival Meetup taken place on Dec. 1st 2022.

The example uses Twitch Gamer Social Network data. It contains Twitch users as nodes and their mutual following relationships as edges.

Task

• Predict how likely users form connection Evaluation Dataset For each player, create a sequence of 5 Positive sample: connected edges, and 5 Negative sampling from unconnected edges. Evaluation Metric NDCG

The example builds a simple GNN with GaphSage layers can be found in src/model.py Data preprocessing and other utils can be found in src/utils.py Model training code can be found in src/train.py

pip install scikit-learn numpy pandas torch tqdm pandarallel

NOTE You will also need to install dgl to construct the graph. Please follow the official instructions to install the version that's compatible to your machine.

cd graph_learning_demo

Download data first.

wget https://snap.stanford.edu/data/twitch_gamers.zip
mkdir data
cd data
unzip ../twitch_gamers.zip

Then run the following

python run.py

You can modify example_user by updaing the input argument for run.run() It will print the top mostly likely users to form a connection with the example user, as well as the least likely users to form a connection with the example user.