📖 Model Card: [REARM]

"[Refining Contrastive Learning and Homography Relations for Multi-Modal Recommendation]",
Shouxing Ma, Yawen Zeng, Shiqing Wu, and Guandong Xu
Published in [ACM MM], 2025.
[Paper Link] [Code Repository]

✨ Overview

• We propose a novel multi-modal contrastive recommendation framework (REARM), which preserves recommendation-relevant modal-shared and valuable modal-unique information through meta-network and orthogonal constraint strategies, respectively.

• We jointly incorporate co-occurrence and similarity graphs of users and items, allowing more effective capturing of the underlying structural patterns and semantic (interest) relationships, thereby enhancing recommendation performance.

• Extensive experiments are conducted on three publicly available datasets to evaluate our proposed method. The experimental results show that our proposed framework outperforms several state-of-the-art recommendation baselines.

---

🧩 Environment Requirement

The code has been tested running under Python 3.6. The required packages are as follows:

• pytorch == 1.13.0
• numpy == 1.24.4
• scipy == 1.10.1

Data

Full data could be downloaded from huggingfac:

• Baby
• Sports
• Clothing

Dataset

We provide three processed datasets: Baby, Sports, and Clothing.

#Dataset	#Interactions	#Users	#Items	Sparsity
Baby	160,792	19,445	7,050	99.88%
Sports	296,337	35,598	18,357	99.96%
Clothing	278,677	39,387	23,033	99.97%

🚀 Example to Run the Codes

The instructions for the commands are clearly stated in the codes.

• Baby dataset

python main.py --dataset='baby'  --num_layer=4   --reg_weight=0.0005 --rank=3  --s_drop=0.4  --m_drop=0.6 --u_mm_image_weight=0.2  --i_mm_image_weight=0  --uu_co_weight=0.4 --ii_co_weight=0.2  --user_knn_k=40  --item_knn_k=10 --n_ii_layers=1 --n_uu_layers=1 --cl_tmp=0.6  --cl_loss_weight=5e-6 --diff_loss_weight=5e-5

• Sports dataset

python main.py --dataset='sports' --num_layer=5  --reg_weight=0.05 --rank=7  --s_drop=1  --m_drop=0.2 --u_mm_image_weight=0  --i_mm_image_weight=0.2  --uu_co_weight=0.9 --ii_co_weight=0.2  --user_knn_k=25  --item_knn_k=5 --n_ii_layers=2 --n_uu_layers=2 --cl_tmp=1.5  --cl_loss_weight=1e-3 --diff_loss_weight=5e-4

• Clothing dataset

python main.py --dataset='clothing' --num_layer=4  --reg_weight=0.00001 --rank=3  --s_drop=0.4  --m_drop=0.1 --u_mm_image_weight=0.1  --i_mm_image_weight=0.1  --uu_co_weight=0.7 --ii_co_weight=0.1  --user_knn_k=45  --item_knn_k=10 --n_ii_layers=1 --n_uu_layers=1 --cl_tmp=0.03  --cl_loss_weight=1e-6 --diff_loss_weight=1e-5

REARM

The released code consists of the following files.

--data
    --baby
    --clothing
    --sports
--utils
    --configurator
    --data_loader
    --evaluator
    --helper
    --logger
    --metrics
    --parser              
--main
--model
--trainer

Citation

If you want to use our codes and datasets in your research, please cite:

@inproceedings{REARM,
  title     = {Refining Contrastive Learning and Homography Relations for Multi-Modal Recommendation,
  author    = {Ma, Shouxing and 
               Zeng, Yawen and 
               Wu, Shiqing and 
               Xu, Guandong},
  booktitle = {Proceedings of the 33th ACM International Conference on Multimedia},
  year      = {2025}
}