This is the raw implementation of our paper Decoding Matters: Addressing Amplification Bias and Homogeneity Issue for LLM-based Recommendation
To reproduce our results, you need to conduct the following pipeline.
# Take the book dataset as an example
# Download the dataset
wget https://datarepo.eng.ucsd.edu/mcauley_group/data/amazon_v2/categoryFiles/Books.json.gz
wget wget https://datarepo.eng.ucsd.edu/mcauley_group/data/amazon_v2/metaFiles2/meta_Books.json.gz
# Unzip
gunzip Books.json.gz
gunzip meta_Books.json.gz
# Preprocess
python ./code/preprocess.py --category "Books"
# Train
bash run.sh # You only need to change the category parameter in script
# Inference and Evaluate
bash evaluate.sh
# Decoding Matters Inference (Our Methods) and Evaluate
bash evaluate2.sh # You need to specify your logits file in the scriptThe results and the parameters of Qwen2-0.5B trained on five Amazon datasets are presented in the following table:
| Dataset | NDCG@10 | HR@10 | Link |
|---|---|---|---|
| CDs_and_Vinyl | 0.077 | 0.109 | link |
| Video_Games | 0.052 | 0.085 | link |
| Toys_and_Games | 0.053 | 0.096 | link |
| Sports_and_Outdoors | 0.099 | 0.120 | link |
| Book | 0.018 | 0.027 | link |
If you're using this code in your research or applications, please cite our paper using this BibTeX:
@article{bao2024decoding,
title={Decoding Matters: Addressing Amplification Bias and Homogeneity Issue for LLM-based Recommendation},
author={Bao, Keqin and Zhang, Jizhi and Zhang, Yang and Huo, Xinyue and Chen, Chong and Feng, Fuli},
journal={arXiv preprint arXiv:2406.14900},
year={2024}
}and
@article{bao2023bi,
title={A bi-step grounding paradigm for large language models in recommendation systems},
author={Bao, Keqin and Zhang, Jizhi and Wang, Wenjie and Zhang, Yang and Yang, Zhengyi and Luo, Yancheng and Chen, Chong and Feng, Fuli and Tian, Qi},
journal={arXiv preprint arXiv:2308.08434},
year={2023}
}