Introduction

The source code of "Preference-Aware Multimodal Contrastive Learning Recommendation Method" (PAMCL).

Environment

The Python and package managers version are:

• Python 3.9.19
• mamba 1.5.8
• conda 24.3.0

Use conda/mamba or pixi (Recommend) to install the environment:

conda env create -f environment.yml
# or
pixi install

How to run

python main.py -c conf/*.yaml

Datasets

Download from google drive.

Pre-trained models

• Vision: openai/clip-vit-base-patch32
• Text: bennegeek/stella_en_1.5B_v5

Warning

Use bennegeek/stella_en_1.5B_v5 to generate text embeddings requires flash-attn which needs a CUDA version of at least 11.6, as indicated by nvcc -V.

Contributions

This project is forked from the original SELFRec. We would like to express our gratitude to Coder-Yu and his team for their outstanding work and open-source contributions.

The main changes are:

• Support multi-modal datasets
• Use Large Language Models to enhance presentations
• Update some deprecated APIs
• Improve code readability by type hint and docstrings
• Refactor some modules

Citation

@ARTICLE{11415690,
  author={Wang, Hairong and Yi, Zhihang and Xu, Zhaojing and Wang, Jing and Li, Hongying},
  journal={IEEE Transactions on Big Data}, 
  title={Preference-Aware Multimodal Contrastive Learning Recommendation}, 
  year={2026},
  volume={},
  number={},
  pages={1-12},
  keywords={Contrastive learning;Feature extraction;Semantics;Recommender systems;User preference;Large language models;Graph neural networks;Computational modeling;History;Computational efficiency;Multimodal recommendation;contrastive learning;large language model;graph neural network;negative sampling},
  doi={10.1109/TBDATA.2026.3668549}}