M³Rec: Selective State Space Models with Mixture-of-Modality Experts for Multi-Modal Sequential Recommendation

This is our Pytorch implementation for ICASSP 2025 paper M³Rec: Selective State Space Models with Mixture-of-Modality Experts for Multi-Modal Sequential Recommendation.

🚀🚀M³Rec is a new multimodal sequential recommendation framework that integrates a Mamba-based selective state space model with Mixture-of-Modality Experts.M³Rec strengthens the modeling of user action sequence dependencies through shared Mamba blocks across modalities and employs modality experts to extract modality-specific user preferences.

Dependencies

All experiments are conducted on an NVIDIA 24GB 3090 GPU. The required packages are as follows:

• Python 3.7+
• PyTorch 1.12+
• CUDA 11.6+
• Install RecBole:
  • pip install recbole
• Install causal Conv1d and the core Mamba package:
  • pip install causal-conv1d>=1.2.0
  • pip install mamba-ssm

Datasets

📢📢We provide public three pre-processed datasets requiring no additional processing on Google Drive. You can download and place them in the "./dataset".

✨✨To facilitate research in multimodal sequential recommendation systems, we provide code for preprocessing the original Amazon datasets in "preprocess/data_preprocess.ipynb". If you find this resource useful for your work, please kindly cite our work.

Run

python run_M3Rec.py

Note

🔥🔥In our experiments, we observed significant performance differences across different python environments. Through detailed analysis, we analyze these discrepancies from version differences between the Mamba's required dependency "causal-conv1d" and "mamba-ssm". To ensure transparency and reproducibility, we have released the training logs under the "./best_log" and corresponding model weights in Google Drive.

Citation

If you find this work helpful to your research, please kindly consider citing our paper.

@inproceedings{guo2025m,
  title={M 3 Rec: Selective State Space Models with Mixture-of-Modality Experts for Multi-Modal Sequential Recommendation},
  author={Guo, Xu and Zhang, Tong and Xue, Yufei and Wang, Chenxu and Wang, Fuyun and Cui, Zhen},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

Acknowledgment

This project is based on Mamba4Rec, Mamba, Causal-Conv1d, and RecBole. Thanks for their excellent works.