Sociologically-Informed Neural Network (SINN)

Source code of our KDD 2022 paper: Predicting Opinion Dynamics via Sociologically-Informed Neural Networks [Link to the paper]

Requirements

• Python 3.9
• numpy==1.19.5
• opencv_python==4.5.2.52
• pandas==1.2.0
• python_dateutil==2.8.2
• pytreebank==0.2.7
• scikit_learn==1.0.2
• scipy==1.6.0
• torch==1.10.0
• torchmetrics==0.5.1
• torchtext==0.11.0
• torchvision==0.9.1
• transformers==4.8.2

Directory description

• src/: The models are stored here.

  • voter.py: PyTorch implementation of Voter model.

  • degroot.py: PyTorch implementation of DeGroot model.

  • aslm.py: PyTorch implementation of asynchronous linear model (AsLM).

  • slant.py: PyTorch implementation of a nonlinear generative model for opinion dynamics (SLANT).

  • slant_plus.py: PyTorch implementation of SLANT+.

  • nn.py: PyTorch implementation of neural network (NN).

  • sinn.py: PyTorch implementation of the proposed method, Sociologically-Informed Neural Networks (SINN).

• working/: The preprocessed datasets are stored here.

  • posts_final_sample_twitter_Abortion.tsv: Dataset of 100 most active users in Twitter Abortion dataset.

    Each row has the format: {user_id},{opinion},{time}.

    {opinion} is manually annotated class label.

    Due to the privacy concern, we removed tweet text.

  • posts_final_synthetic_consensus.tsv: Synthetic dataset generated using SBCM (stochastic opinion dynamics model) with exponent parameter $\rho=-1.0$.

  • posts_final_synthetic_clustering.tsv: Synthetic dataset generated using SBCM with exponent parameter $\rho=0.1$.

  • posts_final_synthetic_polarization.tsv: Synthetic dataset generated using SBCM with exponent parameter $\rho=1.0$.

    Each row has the format: {user_id},{opinion},{time}.

Data Generation & Collection

Data generation

We provide synthetic datasets in working/. But you can also generate these datasets.

• Generate three synthetic datasets

  python3 simulate.py
  

Data collection

We provide sample real data in working/. But you can also query the dataset via API.

1. Set your Bearer Token in your environment variable by running the following command in your terminal.

   export BEARER_TOKEN='<your-bearer-token>'
   

2. Get tweets by query search by running

   python3 collect_twitter.py
   

   Get twitter probiles by username

   python3 collect_twitter_user.py  
   

   Get Reddit posts

   python3 collect_reddit.py
   

Data preprocessing

• Real-world datasets preprocessing

  python3 preprocess.py 
  

• Change data formats

  python3 convert_data.py 
  

How To Use

Example of Usage

• Training: Run main_sinn.py file to train and evaluate the proposed method with default settings.

  python3 main_sinn.py 
  

Advanced Usage

• To train all models, run the bash shell script

  ./run_all_methods.sh
  

• To tune hyperparameters in NN and our SINN, run the bash shell script

  ./run_grid_search.sh
  

• To specify the parameters, run

python3 main_sinn.py \
   --method SINN \
   --dataset synthetic_consensus \
   --save_dir output/ \
   --num_hidden_layers 5 \
   --hidden_features 8 \
   --alpha 0.1 \
   --beta 0.1 \
   --num_epochs 500 \
   --lr 0.001 \
   --K 1 \
   --type_odm SBCM

• method: str (default=SINN)

  Options are "Voter", "DeGroot", "AsLM", "SLANT", "SLANT+", "NN", "SINN"

• dataset: str (default=synthetic_consensus)

  Options are "synthetic_consensus", "synthetic_clustering", "synthetic_polarization", "sample_twitter_Abortion"

• save_dir specifies the path to save the trained model. The model path defaults to be "./output" if not specified.

• num_hidden_layers specifies the number of layers $L$ in the neural network.

• hidden_features specifies the number of units per layer $N_u$ in the neural network.

• alpha specifies the trade-off hyperparameter $\alpha$.

• beta specifies the trade-off hyperparameter $\beta$.

• lr specifies learning rate.

• K specifies dimension of latent space.

• type_odm specifies the choice of opinion dynamics model.

• use_profile: bool (default=False)

  If True, then use profile descriptions of Twitter users as input of the neural network.

• To visualize evaluation results, run plotting.py file.

  python3 plotting.py 
  

Citation

If you use this code or data, please cite the following paper:

@inproceedings{okawa2022sinn,
  author    = {Okawa, Maya and Iwata, Tomoharu},
  title     = {Predicting Opinion Dynamics via Sociologically-Informed Neural Networks},
  booktitle = {{KDD} '22: The 28th {ACM} {SIGKDD} Conference on Knowledge Discovery and Data Mining},
  pages     = {},
  year      = {2022}
}