Repository of the AAAI Submission "IOHunter: Graph Foundation Model to Uncover Online Information Operations".

Reproducibility Steps

Data preprocessing

• Clone the repository in your local space
• Download the data from this zenodo public link and unzip it in the main folder.
  • Your project tree should resemble this structure:
    • /src
    • /data/
    • /data/processed/UAE
    • /data/processed/cuba
    • /data/processed/russia
    • /data/processed/venezuela
    • /data/processed/iran
    • /data/processed/china

Running scripts

• Each running script takes as input several parameters, a typical run is the following:
  • python run_MultiModalGNN_CrossAttention.py --dataset russia --lr 1e-2 --early 30 --gnn sage
• Argument dataset accepts values in UAE, cuba, russia, venezuela, iran, china (same dataset names as in the paper).
• Argument lr accepts continuous values and it represents the learning rate of the Adam optimizer.
• Argument early is the number of epochs without improvement in Macro-F1 after which the early stopping halts the training.
• Argument gnn accepts values in gcn, sage and represents whether the backbone GNN model is a GCN or a Sage.
• You can also add the argument undersampling to specify whether you want to train the model in a data scarcity regimes. It accepts values in 0.5, 0.75, 0.9, 0.95, 0.99, 0.999 as used in the paper.