DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025)

This repository contains the official implementation for the paper "DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection", accepted by the AAAI Conference on Artificial Intelligence (AAAI) 2025.

Paper

DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection Authors: Weihai Lu, Yu Tong, Zhiqiu Ye Conference: AAAI 2025 Link: https://ojs.aaai.org/index.php/AAAI/article/view/32036

Requirements

The required packages are listed in requirements.txt. The setup is based on the environment used for MMDFND.

You can install the dependencies using pip:

pip install -r requirements.txt

(Note: Please ensure you have compatible versions of PyTorch, CUDA, etc., installed based on your system setup and the potential implicit dependencies from MMDFND's original requirements.)

Directory Structure

The project is organized as follows:

.
├── Weibo_21                # Processed data for Weibo_21 dataset
│   ├── train_clip_loader.pkl
│   ├── train_loader.pkl
│   ├── train_origin.csv
│   └── ...
├── data                    # Processed data for the primary/weibo dataset 
│   ├── train_clip_loader.pkl
│   ├── train_loader.pkl
│   ├── train_origin.csv
│   └── ...
├── src                     # Source code directory
│   ├── clip_data_pre.py    # CLIP feature preprocessing script
│   ├── data_pre.py         # General data preprocessing script
│   ├── main.py             # Main training and evaluation script
│   ├── model               # Model definitions
│   │   ├── dammfnd.py      # DAMMFND model implementation
│   │   ├── layers.py       # Custom layers
│   │   └── pivot.py        # Pivot mechanism implementation
│   ├── models_mae.py       # MAE model components (if used)
│   ├── param_model         # Saved model parameters/checkpoints
│   │   └── dammfnd
│   ├── pretrained_model    # Directory for downloaded pretrained models
│   │   ├── chinese_roberta_wwm_base_ext_pytorch # Example: RoBERTa model
│   │   └── w2v             # Example: Word2Vec model
│   ├── run.py              # Potential helper script to run experiments
│   ├── util                # Utility functions (MAE related)
│   │   ├── __pycache__
│   │   ├── crop.py
│   │   ├── datasets.py
│   │   ├── lars.py
│   │   ├── lr_decay.py
│   │   ├── lr_sched.py
│   │   ├── misc.py
│   │   └── pos_embed.py
│   ├── utils               # General utility functions and data loaders
│   │   ├── __pycache__
│   │   ├── clip9.py
│   │   ├── clip_dataloader.py
│   │   ├── dataloader.py
│   │   ├── utils.py
│   │   └── weibo21_clip_dataloader.py
│   ├── weibo21_clip_data_pre.py # CLIP preprocessing specific to Weibo_21
│   └── weibo21_data_pre.py    # General preprocessing specific to Weibo_21
├── weibo                   # Raw data files for Weibo dataset (CSV format)
│   ├── test_2_domain.csv
│   ├── train_2_domain.csv
│   └── val_2_domain.csv
└── weibo21                 # Raw data files for Weibo_21 dataset (XLSX format)
    ├── test_2_domain.xlsx
    ├── train_2_domain.xlsx
    └── val_2_domain.xlsx
├── requirements.txt        # Python package dependencies
└── README.md               # This file

Data Preparation

Please follow the specific instructions from the MMDFND repository's Data Preparation section for acquiring and initially placing the raw data.

Note: If the data linked or described by the MMDFND authors is unavailable, please contact [email protected] to request the complete dataset.

Pretrained Models

This project utilizes pretrained models like RoBERTa and potentially Word2Vec or CLIP encoders.

1. Download Models: Please refer to the MMDFND repository's Pretrained Models section for instructions on downloading the required models.
2. Place Models: Store the downloaded pretrained model files within the src/pretrained_model/ directory, following the sub-directory structure shown (e.g., src/pretrained_model/chinese_roberta_wwm_base_ext_pytorch/).

Training

To train the DAMMFND model:

1. Navigate to the source code directory:

   cd src

2. Run the main script:

   python main.py

You can switch between datasets (e.g., the default one using data/ or the Weibo_21 dataset using Weibo_21/) by modifying the relevant dataset path configurations within the main.py script or potentially through command-line arguments if implemented. Please inspect main.py for details on dataset selection and other hyperparameters.

Acknowledgements

This codebase is developed based on the implementation of MMDFND. We sincerely thank the original authors for sharing their code and groundwork.

Log

Based on the prepared code, experiments were conducted on the weibo and weibo21 datasets. Refer to src/train_weibo.log and src/train_weibo21.log for the respective logs.

Note: Slight discrepancies in the results might be observed across different hardware configurations.

Citation

If you find this work useful for your research, please consider citing our paper:

Lu W, Tong Y, Ye Z. DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(1): 559-567.

Or in BibTeX format:

@inproceedings{lu2025dammfnd,
  title={DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection},
  author={Lu, Weihai and Tong, Yu and Zhiqiu, Ye},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={1},
  pages={559--567},
  year={2025},
  url={https://ojs.aaai.org/index.php/AAAI/article/view/32036}
}