git clone [email protected]:chaudatascience/fake_news_detection_gnns.git
cd fake_news_detection_gnns
conda env create -f environment.yml
conda activate fake_news
sh run.sh
python -m src.fake_news.fake_news_detection --dataset gossipcop --epochs 300 --early_stopping 50 --batch 128
Arguments
'--dataset', type=str, default="gossipcop",
help="name of the dataset, either 'gossipcop' or 'politifact' "
'--early_stopping', type=int, default=50,
help="stop training after `early_stopping` non-decreasing val loss epochs"
'--cuda', type=str, default="auto",
help="choose x for 'cuda:x', or using most available GPU by default"
'--batch', type=int, default=128,
help="batch size for training"
'--lr', type=float, default=0.001,
help="learning rate"
'--weight_decay', type=float, default=0.01,
help="weight decay for lr"
'--epochs', type=int, default=300,
help="training epochs"
'--dropout', type=float, default=0,
help="dropout"
'--pooling', type=str, default="global_max_pool",
help="one of [global_mean_pool, global_max_pool, global_attention, global_attention_with_relu, global_attention_with_relu_linear]"
'--gat_layer', type=str, default="GATConv",
help="one of ['OurGATNet', GATConv', 'GATv2Conv', 'SuperGATConv']"
'--hid_dims', type=List[int], default=[32],
help="hidden dimensions for GATs"
'--news_dim', type=int, default=64,
help="dimensions for news"
'--readout_dim', type=int, default=64,
help="dimensions for graph readout"
'--num_heads', type=int, default=1,
help="num attention heads for each GAT layer"
'--feature', type=str, default='content',
help="feature type: [profile, spacy, bert, content],`content` means 300-d word2vec+10-d profile"
'--only_gat', type=bool, default=False,
help="Only use GAT (testing purpose)"
For hyper-param tuning, use --hyper_param flag, and use --config_file to point to the yaml config file.
Example of the config file can be found here.
python -m src.fake_news.fake_news_detection --hyper_param --config_file demo
Some important files
- fake_news_detection.py: Main file for fake news detection
- fake_news_net.py: Our implementation for Fake News Net
- GAT package: Our implementation of Graph Attention Networks (GATs)
- GAT package v2: Another version of our implementation of GATs.
Datasets from UPFD (SIGIR'21) paper.
Fake and Real news on Twitter verified by
- Politifact: Journalists and domain experts review the political news and provide fact-checking evaluation results.
- Gossipcop: GossipCop provides rating scores on the scale of 0 to 10 to classify a news story.
Train and val losses on Gossipcop dataset
Resulting table
Analyse Results: More details of the results (e.g., train and val losses, accuracy by number of attention heads) can be found at this Jupyter notebooks
Chau Pham, Sheng Huang, and Efe Sencan
| Title | Tag | Conference | Links | Notes |
|---|---|---|---|---|
| Semi-Supervised Classification with Graph Convolutional Networks | GCN | 2016 | Paper Video | |
| Inductive Representation Learning on Large Graphs | Graph SAGE | 2017 | Paper Video | |
| Graph Attention Networks | GAT | ICLR 2018 | Paper Video DGLblog MainCode Code2 Code3 | |
| A Generalization of Transformer Networks to Graphs | Graph Transformers | AAAI 2021 | Paper Code | General form of GAT |
| Understanding Convolutions on Graphs | GNN tutorials | Distill 2021 | part2 part1 | |
| Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges | small book + tutorials on GNNs | 2021 | Paper Tutorials ICLR2021 MLStreetTalk | |
| Attention is all you need | Transformer | Neurips 2017 | Paper video | |
| Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention | Linear Transformer | ICML 2020 | Paper video | Reduce time complexity O(N^2) to O(N) |
| GNN papers + DGL examples | DGL | Tutorials | ||
| User Preference-aware Fake News Detection | UPFD | SIGIR 2021 | Paper Code | datasets, our model was based on this awesome paper |
To compare our own GAT implementation with the original GAT implementation, and compare our fake news net with UPFD, we borrow official implentations from https://github.com/PetarV-/GAT and https://github.com/pyg-team/pytorch_geometric/
-
paper_gat.py are from https://github.com/PetarV-/GAT
-
upfd.py: UPFD implementation
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent