MSCA (WWW'26)

PyTorch implementation for MSCA proposed in the following paper:

Multi-view Semantic Contrastive Alignment for Multimodal Recommendation
Jiuqiang Li, Hongjun Wang*
In WWW 2026
Paper

Overview

Environment

• Python 3.8.10
• PyTorch 1.11.0+cu113

For dependency details, refer to requirements.txt.

Dataset

Download from Google Drive: Baby/Sports/Electronics (Raw Data). The data includes image and text features provided by the MMRec framework, extracted from VGG and Sentence-Transformers. Preprocessing from raw data can be found here.

Download a supplementary dataset for micro-video recommendation: MicroLens (Raw Data) within MMRec.

Training and Evaluation

1. Download the datasets and place them in the data folder.

2. Set the hyperparameters in the src/configs/model/MSCA.yaml file.

3. Run:

cd ./src
python main.py -m MSCA -d {dataset_name}

4. Test:

python test.py -m MSCA -d {dataset_name} -c {checkpoint_path}

Performance Comparison

Reproducibility

We report the best hyperparameters of MSCA to reproduce the results in Table 2 and 6 of our paper.

Dataset	n_layers	fusion_coeff	cl_weight	reg_weight
Baby	2	0.4	0.005	3e-7
Sports	3	0.3	0.005	5e-8
Electronics	4	0.2	0.01	5e-10
MicroLens	4	0.3	0.01	5e-9

The training logs and model checkpoints are provided below:

Dataset	Download
Baby	log	checkpoint
Sports	log	checkpoint
Electronics	log	checkpoint
MicroLens	log	checkpoint

Citation

If you find MSCA helpful to your research, please consider citing the following paper.

@inproceedings{li2026multi,
  title={Multi-view Semantic Contrastive Alignment for Multimodal Recommendation},
  author={Li, Jiuqiang and Wang, Hongjun},
  booktitle={Proceedings of the ACM Web Conference 2026},
  pages={5941--5952},
  year={2026}
}

Licensed under the GNU GPL v3.0. See LICENSE.

Acknowledgement

​​This repository is based on MMRec. Thanks for their work.