---
license: apache-2.0
pipeline_tag: any-to-any
library_name: fuxictr
---

## Official implementation code of DISCO.AHU team

## 🔥 Winning 2nd Place in WWW2025 Multimodal CTR Prediction Challenge Track

This repository contains the model as presented in [Quadratic Interest Network for Multimodal Click-Through Rate Prediction](https://huggingface.co/papers/2504.17699).

### 🔥 Follow to perfectly reproduce the results of this code.

- To facilitate reproducibility, we share the model code on GitHub: https://github.com/salmon1802/QIN
- 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

### Data Preparation

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

2. Unzip the data files to the `data` directory

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

### Environment

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
```

### How to Run

 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.

#### Tips
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.