The source code for SIGIR 2022 paper: "Unsupervised Belief Representation Learning with Information-Theoretic Variational Graph Auto-Encoders"

To run InfoVGAE on Eurovision dataset:

python3 main.py --config_name InfoVGAE_eurovision_3D

To run InfoVGAE on Election dataset:

python3 main.py --config_name InfoVGAE_election_3D

To run InfoVGAE on Voteview 105th Congress dataset:

python3 main.py --config_name InfoVGAE_bill_3D

To run InfoVGAE on TIMME dataset:

python3 main.py --config_name InfoVGAE_timme_3D

To run InfoVGAE on TIMME dataset with follow (friend) links:

python3 main.py --config_name InfoVGAE_timme_follow_3D

The embeddings, labels, figures, and top-k tweets (only applicable for Twitter datasets), etc, will be saved in ./output

We uploaded the pre-processed datasets with smaller size, due to the file size limits of Github. The datasets are located in dataset/election, dataset/eurovision, and dataset/bill. It may takes some time to clone this repo (297MB). After cloning this repo, please run:

unzip dataset/bill/bmap2.pkl.zip; unzip dataset/bill/data_80_115.pkl.zip

Evaluation will be automaticly triggered after the training process. To evaluate again, modify the evaluator.init_from_dir() in evaluate.py.

General

--use_cuda: training with GPU

--epochs: iterations for training

--learning_rate: learning rate for training

--device: which gpu to use. empty for cpu.

--num_process: num process for pandas processing

Data

--data_path: csv path for data file

--stopword_path: stopword path for text parsing

--kthreshold: tweet count threshold to filter not popular tweets.

--uthreshold: user count threshold to filter not popular users.

For InfoVGAE model

--hidden1_dim: the latent space dimension of first layer

--hidden2_dim: the latent space dimension of target layer

Result

--output_path path to save the result