TEARS

Text Representations for Scrutable Recommendations (link to come)

Data

We provide detailed instructions on how to reproduce the datasets used in the data directory

User Summaries

We provide user summaries in ./saved_user_summaries. These can be remade with

python -m preprocess.make_in_context_data --data_name=${data_name}

To simply use the preprocessed data directly first run:

tar xvzf data_preprocessed.tar.gz
tar xvzf saved_user_summary.tar.gz

tar xvzf data.tar.gz contains all original data, which can be redownloaded using the instructions provided in data, but we include here for convinience.

tar xvzf data_preprocessed.tar.gz contains all relevant processed data that is needed to run the code.

saved_user_summary.tar.gz contains all LLaMA and GPT user summaries, as well as augmented summaries used to evaluate large-scope and fine-grained controllability.

Execution

We provide an example execution command below, but note other examples given in /execute

We implement EASE,MVAE,MDAE, MacridVAE,RecVAE and their appropriate TEARS models

To execute the training pipeline run:

python -m trainer.train \
    --seed=2024 \
    --embedding_module=${module} \
    --data_name=${dataset} \
    --dropout=.4 \
    --epochs=30 \
    --lora_alpha=16 \
    --lora_r=64 \
    --bs=64 \
    --lr=.0001 \
    --scheduler=None \
    --epsilon=.5 \
    --eval_control \
    --wandb

Requirements

Acknowledgements

We thank the authors of the following repositories for their useful codebases that where a key role in the execution of this project:

MVAE/MDAE https://github.com/younggyoseo/vae-cf-pytorch

RecVAE https://github.com/ilya-shenbin/RecVAE

MacridVAE (adapted to torch from) https://jianxinma.github.io/disentangle-recsys.html