DisCo

This is the implementation of "DisCo: Towards Harmonious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation".

Data

The raw data of the three dataset is available at:

https://grouplens.org/datasets/movielens/1m/

https://cseweb.ucsd.edu/jmcauley/datasets.html

https://files.grouplens.org/datasets/mov

To generate the processed data, run python data_preprocess.py in dissem_preprocess folder.

You can use an arbitrary LLM model to generate semantic embedding, for ours, we use vicuna-13b:

https://vicuna.lmsys.org/

Baseline

run the run_base.py file in run folder, we provide three methods to choose different ways to embed user history.

• No history, use --model={model}, for example python run_base.py --model=DeepFM
• Average pooling on user history, use --model={model}Mean, for example python run_base.py --model=DeepFMMean. This is our baseline model in main results table
• Attention on user history, use --model={model}Att, for example python run_base.py --model=DeepFMAtt.

Our method

run the run_DisCo.py file in run folder, use --model=DisCo{model} to specify the backbone.

Citation

If you find this repo useful, please cite our paper.

@misc{du2024disco, title={DisCo: Towards Harmonious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation}, author={Kounianhua Du and Jizheng Chen and Jianghao Lin and Yunjia Xi and Hangyu Wang and Xinyi Dai and Bo Chen and Ruiming Tang and Weinan Zhang}, year={2024}, eprint={2406.00011}, archivePrefix={arXiv}, primaryClass={cs.IR} }