Fake News Propagation

Project idea and implementation is based on the ICWSM 2020 paper "Hierarchical Propagation Networks for Fake News Detection: Investigation and Exploitation" (https://arxiv.org/abs/1903.09196)

Dataset

Hierarchical propagation networks are constructed using the news samples from the FakeNewsNet dataset. To adhere to Twitter's privacy policy, the user information is anonymized and tweet contents are not shared.

The dataset is formatted in the networkx graph JSON format and available at data directory, categorized based on the news source (Politifact/GossipCop) and label.

Each node in the graph contains the random tweet id, exact epoch timestamp, node type, random user id attributes. Additionally, the nodes in the retweet networks(macro networks) have bot score attribute and the nodes in the reply chain network(micro-network) have the sentiment of the tweet content.

For the results mentioned in the paper, randomly down-sampled news ids provided in the data/sample_ids is used.

To Run:

• To use this dataset, un-zip the file nx_network_data.zip file in data directory and one can use an example in load_dataset.py to load the dataset.
• Install the dependencies in requirements.txt first
• To extract features and run the model, use basic_model.py's main function or run FN.ipynb cells.
• Function get_classificaton_results_tpnf_by_time in basic_model can be used to prune the graphs by time and work on the pruned dataset.
• To explore the GCN model and check results use GCN_new.ipynb and run the cells