Fully Test-time Adaptation for Tabular Data

🏠 Homepage • 📃 Paper

This is the official code for paper titled: Fully Test-time Adaptation for Tabular Data

Quick Start

1. Prepare Python Environment

Clone FTTA repository, create conda environment, and install required packages. This repository is constructed based on Tableshift repository. For any detail of TableShift, please visit TableShift at tableshift.org

NOTE: The requirements.txt file in FTTA is silently different from that in TableShift repo due to the package compatibility issues in Python.

git clone https://github.com/WNJXYK/FTTA.git
cd FTTA/src
conda env create -f environment.yml
conda activate ftta
python examples/run_expt.py

The final line above will print some detailed logging output as the script executes. When you see training completed! test accuracy: 0.6221 your environment is ready to go! (Accuracy may vary slightly due to randomness.)

2. Run FTTA Approach

Run the FTTA approach using the command below:

conda activate ftta
python model_train.py --experiment {exp_name} --model {model_name}

For example, run the command below, soon it will show the result of FTTA methods and Unadapt methods.

conda activate ftta
python model_train.py --experiment diabetes_readmission --model mlp

For the first time running, dataset will be automatically download to tmp folder. For better reproduction our results, we provide pre-trained models for supported experiment and models, each experiment setting provides one model in scr/models folder. Supported experiment and model are showed follows, we will expand the support list in the future.

Supported models and experiments in FTTA is

Dataset	String Identifier
Voting	anes
ASSISTments	assistments
HELOC	heloc
Hospital Readmission	diabetes_readmission

Models	String Identifier
MLP	mlp
TabTransformer	tabtransformer
FT-Transformer	ft_transformer

3. Explore more on FTTA methods

If you want to explore FTTA on more datasets, some key files may be helpful to you.

Some key code files

Path	Function
/tableshift/FTTA_src/FTTA.py	Definition of FTTA methods. When creating a FTTA class, a well trained tabular model, prior of training set and a optimizer type is need specifying.
/tableshift/models/torchutils.py	The evaluation process is defined here. Tf you want to test some other methods, modify evaluate function is needed.
/tableshift/models/default_hparams.py	Algorithms' default hyper-parameters is defined here. We set 'n_epoch = 0' for test-time adaptation using trained model or provided model. If you want to train your own model, set 'n_epoch' to a positive integer.

4. Dataset Availability

The availability of dataset is the same as TableShift benchmark. If you want to add more dataset, see the guidelines of TableShift benchmark at tableshift.org.

5. Q&A

If you have any questions, feel free to contact us at [email protected] or submit an issue here.

Acknowledgements

We thank the author of TableShift benchmark providing a convenient framework to develop tabular algorithms.

Citation

Please cite the paper if you refer to our code or paper from FTTA.

@inproceedings{zhou24ftta,
    author       = {Zhi Zhou and Kun-Yang Yu and Lan-Zhe Guo and Yu-Feng Li},
    title        = {Fully Test-time Adaptation for Tabular Data},
    booktitle    = {Proceedings of the 39th AAAI conference on Artificial Intelligence},
    year         = {2025}
}