RecGRELA

This repository contains the reference code for the paper Gated Rotary-Enhanced Linear Attention for Long-term Sequential Recommendation. The paper is released on arXiv.

1. Overall

2. Requirements

Here are our main environment dependencies for running the repository：

• NVIDIA-SMI 535.183.01
• cuda 12.2
• python 3.11.5
• pytorch 2.4.0
• recbole 1.2.0
• casual-conv1d 1.4.0
• timm 1.0.11

3. Datasets

This repository contains the ML-1M dataset. If you want to train our model on other datasets, the ML-20M, ML-32M, and Netflix datasets can be downloaded from Google Drive. ML-32M can also be found at MovieLens and processed by conversion tools.

4. Run

To reproduce the results reported in our paper, just run it:

python run_RecGRELA.py

5. Results

You can also check the training log in📁 log.

6. References

If you find this code useful or use the toolkit in your work, please consider citing:

@article{hu2025gatedrotaryenhancedlinearattention,
      title={Gated Rotary-Enhanced Linear Attention for Long-term Sequential Recommendation}, 
      author={Juntao Hu and Wei Zhou and Huayi Shen and Xiao Du and Jie Liao and Junhao Wen and Min Gao},
      journal={arXiv preprint arXiv:2506.13315},
      year={2025}
}

Acknowledgment

Our code references RecBole, Mamba4Rec, and Causal-Conv1d. We appreciate their outstanding work and open source.