AtCAF

The code is related to the paper: AtCAF: Attention-based causality-aware fusion network for multimodal sentiment analysis.

Datasets

You can download the CMU-MOSI and CMU-MOSEI datasets using CMU-MultimodalDataSDK.

You can download the UR-FUNNY dataset using UR-FUNNY resp.

You can download the ood version of CMU-MOSI and CMU-MOSEI datasets in https://msa-clue.wixsite.com/clue.

Preparation

Create a folder named npy_folder in the root directory.

mkdir npy_folder

In order to perfectly replicate the precision, please use these functions from model.py to generate the global dictionary initialization as an .npy file and place it in the npy_folder.

gen_npy(enc_word.mean(dim=1).cpu(), self.hp.dataset, n_clusters=25)
gen_npy(enc_word.mean(dim=1).cpu(), self.hp.dataset, n_clusters=50) 
gen_npy(enc_word.mean(dim=1).cpu(), self.hp.dataset, n_clusters=100)  
gen_npy(enc_word.mean(dim=1).cpu(), self.hp.dataset, n_clusters=200)  

Training

python main.py

Environment Requirements

python == 3.8.8

torch == 1.8.1

numpy == 1.20.0

Citation

If you use this code please cite it as:

@article{huang2025atcaf,
  title={AtCAF: Attention-based causality-aware fusion network for multimodal sentiment analysis},
  author={Huang, Changqin and Chen, Jili and Huang, Qionghao and Wang, Shijin and Tu, Yaxin and Huang, Xiaodi},
  journal={Information Fusion},
  volume={114},
  pages={102725},
  year={2025},
  publisher={Elsevier}
}

Thank you for your support. If you have any questions, feel free to post them in the issues or contact us via irelia@zjnu.edu.cn.