ISOLATOR

This repository is the official implementation of the paper:
"Unveiling the Impact of Multi-modal Content in Multi-modal Recommender Systems"
Accepted at ACM Multimedia 2025.

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

@inproceedings{10.1145/3746027.3755300,
  author    = {Xv, Guipeng and Li, Xinyu and Liu, Yi and Lin, Chen and Wang, Xiaoli},
  title     = {Unveiling the Impact of Multi-modal Content in Multi-modal Recommender Systems},
  year      = {2025},
  isbn      = {9798400720352},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  url       = {https://doi.org/10.1145/3746027.3755300},
  doi       = {10.1145/3746027.3755300},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  pages     = {6093–6102},
  numpages  = {10},
  keywords  = {causal inference, multi-modal recommender system, user-side content bias},
  location  = {Dublin, Ireland},
  series    = {MM '25}
}

Environment

The code has been tested with the following dependencies:

• Python == 3.10.13
• PyTorch == 2.0.0
• NumPy == 1.23.5
• Pandas == 1.5.3
• PyYAML == 6.0
• SciPy == 1.11.4

Quick Start

1. Prepare the data
   Download the datasets and place them under the ./data/ directory. Then generate the required multi-modal embeddings as described in the paper.

2. Choose the variant

   • src_D/ corresponds to D-ISOLATOR (Debiasing Item-side Content)
   • src_U/ corresponds to U-ISOLATOR (Debiasing User-side Content)
     Navigate to the appropriate directory and modify the hyperparameter search space in the configuration file (config/*.yaml) if needed.

3. Run training and evaluation
   Execute the following command:

   python -u main.py --model=VBPR --dataset=baby --gpu_id=0

   Replace VBPR with the desired backbone model and baby with your dataset name. The -u flag ensures unbuffered output.

Notes

• For more details on the method, datasets, and hyperparameter settings, please refer to our paper.
• If you encounter any issues or have questions, feel free to open an issue on GitHub.