It is a novel generic and efficient sequential recommendation backbone, which explores the seamless adaptation of Mamba for sequential recommendations. Specifically, SSD4Rec marks the variable- and long-length item sequences with sequence registers and processes the item representations with bidirectional Structured State Space Duality (SSD) blocks. This not only allows for hardware-aware matrix multiplication but also empowers outstanding capabilities in variable-length and long-range sequence modeling.
SSD4Rec: A Structured State Space Duality Model for Efficient Sequential Recommendation
Haohao Qu$$\dagger$$ , Yifeng Zhang$$\dagger$$ , Liangbo Ning, Wenqi Fan*, Qing Li.
$$\dagger$$ The authors contribute equally to this paper.
* Corresponding author.
Paper: https://arxiv.org/abs/2409.01192
You can refer to the required environment specifications in environment.yaml.
An simple example to run SSD4Rec on the ML1M (Default) dataset:
python main.pyWe provide four processed datasets: ML1M, Amazon-Beauty, Amazon-Games, and KuaiRand-Pure. If you want to run experiments on other datasets, you should go to config.yaml and replace the variable of dataset correspondly.
@article{qu2024ssd4rec,
title={Ssd4rec: a structured state space duality model for efficient sequential recommendation},
author={Qu, Haohao and Zhang, Yifeng and Ning, Liangbo and Fan, Wenqi and Li, Qing},
journal={arXiv preprint arXiv:2409.01192},
year={2024}
}This project is based on Mamba and RecBole. Thanks for their excellent works.
More updates will be posed in the near future! Thank you for your interest.