This is the implementation and experiment results of the AAAI 2023 oral paper T2G-FORMER: Organizing Tabular Features into Relation Graphs Promotes Heterogeneous Feature Interaction (see Arxiv version for the full appendix)

Feel free to report any issues or questions.

The implementation of T2G-Former in the original paper is bin/t2g_former.py.

Here are some directories for experiment results.

• results: results of main experiment, including scores and running configs of all models on all datasets. Here are some important keys in experiment output (json file).
  • final: final test score after finetune.
  • best: best test score during finetune.
  • cfg: model configuration used for the experiment.
• configs: tuned configuration of each model on each dataset.
• ablations: coming soon ...
• figures: coming soon ...

• Basic packages: requirements.txt contains some basic packages to execute the script, some package versions are strict and will be told in the following instructions.

• TabNet (TF version): experiment follows the TabNet version in original paper, if you want to use script for TabNet here, tensorflow 1.x is needed. We execute TabNet with tensorflow-gpu==1.14.0 on GeForce RTX 2080 Ti.

• XGBoost: We execute XGBoost with xgboost==1.7.2. When you encounter problems with script for XGBoost, consider to downgrade the version, it depends on its basic API.

• Other pytorch models: We execute other models with torch==1.9.0. Any version >= 1.9.0 is executable.

  • DANets: need tensorboard >= 1.15.0

• We execute all models (except TabNet TF) on NVIDIA GeForce RTX 3090.

LICENSE: by downloading our dataset you accept licenses of all its components. We do not impose any new restrictions in addition to those licenses. You can find the list of sources in the section "References" of our paper.

1. Download the data from this link(~240M)
2. Unpack the archive to the data directory: tar -zxvf data-t2g.tar.gz -C ./data (mkdir data if not exist)
3. Configure your global path to the project folder (variable PROJ in lib/env.py)

• run_experiment: contains all scripts for our experiments.
  • run_baselines.sh: scripts to finetune all models on datasets. If tuned config is not given, it will load a default configuration we provided.
  • tune_baselines.sh: scripts to tune all model configurations on datasets.

You can modify the script, and execute it as follows:

CUDA_VISIBLE_DEVICES=XXX bash ./run_experiment/run_baselines.sh

# ignore system out
CUDA_VISIBLE_DEVICES=XXX nohup bash ./run_experiment/run_baselines.sh >/dev/null 2>&1 &

You can test your models by adding them to bin directory and bin/__init__.py. Keep the same API we used in other models, and execute your model with script run_baselines.py.

We sincerely appreciate the benchmark provided by Yura52‘s work1 and work2 for fair comparison and script implementation.

@inproceedings{yan2023t2g,
  title={{T2G-FORMER:} Organizing Tabular Features into Relation Graphs Promotes Heterogeneous Feature Interaction},
  author={Yan, Jiahuan and Chen, Jintai and Wu, Yixuan and Chen, Danny Z. and Wu, Jian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2023}
}