Competition Official Website: https://erel-mir.github.io/challenge/results/

• To facilitate reproducibility, we share the model checkpoints on Hugging Face: https://huggingface.co/salmon1802/QIN/tree/main
• In ./checkpoints and ./submission folders have our run logs and submission files, respectively.
• This submission can be reproduced manually by following the actions below, or by directly using the one-click run script run.sh

1. Download the datasets at: https://recsys.westlake.edu.cn/MicroLens_1M_MMCTR

2. Unzip the data files to the data directory

   cd ./data/
   wget -r -np -nH --cut-dirs=1 http://recsys.westlake.edu.cn/MicroLens_1M_MMCTR/MicroLens_1M_x1/

We run the experiments on RTX 4090 GPU of AutoDL.com

Please set up the environment as follows.

• torch==2.0.0+cu118
• fuxictr==2.3.7

conda create -n fuxictr_www python==3.8
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
source activate fuxictr_www

Train the model on train and validation sets:

```
python run_expid.py
```

The parameters QIN_variety_v9 in ./config/qin_config/model_config.yaml are set to the optimal hyperparameters in the environment described above.

It is worth mentioning that after our tests, we find that although the parameter num_row = 4 achieves the best performance in the above environments, there is training instability in some environments.

When this happens, we suggest that sacrificing some performance in favor of setting num_row = 3 reproduces the results well.

If you find our code helpful for your research, please cite the following paper:

@article{li2025quadratic,
  title={Quadratic Interest Network for Multimodal Click-Through Rate Prediction},
  author={Li, Honghao and Li, Hanwei and Zhang, Jing and Zhang, Yi and Yu, Ziniu and Sang, Lei and Zhang, Yiwen},
  journal={arXiv preprint arXiv:2504.17699},
  year={2025}
}