FakeNewsMap

Our IBM Z Datathon submission.

A proof of concept of a Graph Neural Network for identifying misinformation on social networks, utilising a small reliable dataset.

Our initial implementation only takes into account the thread's structure and metadata.

Our second implementation takes into account all the above and also the tweet text data by tokenising each tweet with sbert.

Run Locally

Clone the project

  git clone https://github.com/16hayhurstm/FakeNewsMap

Install dependencies

  pip install -r requirements.txt

Download the dataset and save it under data/

Run either build_graphs.py (fast but worse) or build_graphs_v2.py (slow but better) to build the initial graph model.

python3 build_graphs_v2.py

Run the corresponding training program

• build_graphs.py -> train.py
• build_graphsv2.py -> train_sbert.py

python3 train_sbert.py

Optionally export the model