Official PyTorch implementation of the ACM MM 2024 paper Attribute-driven Disentangled Representation Learning for Multimodal Recommendation.
Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, Mohan S. Kankanhalli
Many recommendation methods focus on learning robust and independent representations by disentangling the intricate factors within interaction data across various modalities in an unsupervised manner. However, such an approach obscures how specific factors, such as category or brand, influence the final recommendation results, making their effects difficult to interpret and control.
To address this issue, we propose Attribute-Driven Disentangled Representation Learning (AD-DRL), which explicitly incorporates attributes from different modalities into the disentangled representation learning process. By assigning a specific attribute to each factor in multimodal features, AD-DRL disentangles the factors at both the attribute level and the attribute-value level. To obtain robust and independent representations for each factor, we first disentangle features both within and across modalities, and then further enhance representation robustness by fusing multimodal features belonging to the same factor.
Empirical results on three public real-world datasets demonstrate the effectiveness of AD-DRL, as well as its interpretability and controllability.
- Attribute-driven disentangled representation learning for multimodal recommendation
- Explicit modeling of attributes to improve interpretability and controllability
- Processed datasets for Amazon-Baby, Amazon-Sports, and Amazon-ToysGames
- Released checkpoints for all three datasets
Install the required dependencies with:
pip install -r requirements.txtWe provide three processed datasets:
- Amazon-Baby
- Amazon-Sports
- Amazon-ToysGames
- Dataset link: Google Drive
Please download the three datasets and place them in:
AD-DRL/AmazonData
For more details, please refer to:
amazon.py
We provide checkpoints for all three datasets.
- Checkpoint link: Google Drive
Please download the checkpoints and place them in:
AD-DRL/checkpoints
python main.py --mode "train" --dataset "Baby" --attribute_dataset "item_attribute_label" --learning_rate 0.0001 --decay_r 0 --decay_f 1 --decay_a 50 --decay_n 1 --temp 1 --num_neg 4 --gpu "0" --n_factors 4 --emb_dim 128python main.py --mode "train" --dataset "Sports" --attribute_dataset "item_attribute_label" --learning_rate 0.0001 --decay_r 1 --decay_f 5 --decay_a 10 --decay_n 0.01 --temp 1 --num_neg 8 --gpu "0" --n_factors 5 --emb_dim 160python main.py --mode "train" --dataset "ToysGames" --attribute_dataset "item_attribute_label" --learning_rate 0.0001 --decay_r 1 --decay_f 1 --decay_a 10 --decay_n 0.01 --temp 1 --num_neg 8 --gpu "0" --n_factors 5 --emb_dim 160python main.py --mode "test" --dataset "Baby" --attribute_dataset "item_attribute_label" --num_neg 4 --gpu "0" --n_factors 4 --emb_dim 128python main.py --mode "test" --dataset "Sports" --attribute_dataset "item_attribute_label" --num_neg 8 --gpu "1" --n_factors 5 --emb_dim 160python main.py --mode "test" --dataset "ToysGames" --attribute_dataset "item_attribute_label" --num_neg 8 --gpu "0" --n_factors 5 --emb_dim 160If you find this work useful for your research, please cite:
@inproceedings{ADDRL,
author = {Zhenyang Li and
Fan Liu and
Yinwei Wei and
Zhiyong Cheng and
Liqiang Nie and
Mohan S. Kankanhalli},
title = {Attribute-driven Disentangled Representation Learning for Multimodal
Recommendation},
booktitle = {Proceedings of the ACM International Conference on Multimedia},
pages = {9660--9669},
publisher = {ACM},
year = {2024}
}