Bridging LLM Embeddings and VAE Parameters for Disentangled Recommendation

This repository is the official implementation of paper

Nhu-Thuat Tran and Hady W. Lauw. 2026. Bridging LLM Embeddings and VAE Parameters for Disentangled Recommendation. Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'2026), Melbourne, Australia, July 20-24, 2026.

Environments

• Anaconda: 4.12.0
• Python: 3.7.5
• OS: MacOS

Requirements

To create virtual environment

conda create --prefix ./lampvae python=3.7.5 -y

Then activate environment

conda activate ./lampvae

To install requirements

pip install -r requirements.txt

Update

We highly recommend using uv for fast dependency management.

Training and Evaluation

1. Create a YAML config file under configs folder as samples.

2. Prepare run.sh file as example below

python main_lampvae.py --dataset steam --config_file ./configs/LampVAE-steam.yaml --gpu 0

3. To run training and evaluation

bash run.sh

Hyper-parameter tuning

We follow instructions from VALID (https://github.com/PreferredAI/VALID) for hyper-parameter tuning.

Then, we tune the key hyper-parameters in LampVAE

• n_gcn_layer: the number of GCN layers 1, 2, 3
• tau: the temperature in prototype selection 0.1, 0.15, 0.2, 0.25
• tau_dec: the temperature in decoder 0.1, 0.15, 0.2, 0.25.
• total_anneal_steps: the number of annealing steps in VAEs 1k, 5k, 10k

Citation

If you find our work useful for your research, please cite our paper as

@inproceedings{LampVAE,
    author = {Tran, Nhu-Thuat and Lauw, Hady W.},
    title = {Bridging LLM Embeddings and VAE Parameters for Disentangled Recommendation},
    booktitle = {Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
    pages = {1756–1766},
    year = {2026}
}