liumucan / inafen

Codes for An interpretable automated feature engineering framework for logistic regression research

Dependencies

For a straight-forward use of INAFEN, you can install the required libraries from requirements.txt: pip install -r requirements.txt

Dataset

We provide the raw data used in this paper in ./Data/Raw_data. If you want to use your own dataset, you can process it into the required format.

Example Usage

1. To run the baseline models in a specific dataset: python main_baselines.py --dataset_name some_dataset --baseline_model some_baseline_model

   The dataset_name can be chosen from [adult, churn, credit-card-econometrics, credit-Japan, dataset_29_credit-a, dataset_31_credit-g, dataset_37_diabetes, dataset_heart-failure, gmsc, South-German-Credit-Prediction] and some_baseline_model can be chosen from [XGB, RF, LR, DT, KNN, SVM].

   For example, you can run the experiment with logistic regression on the adult dataset with python main_baselines.py --dataset_name adult --baseline_model LR.

   The generated results can be found in ./Results/dataset_name/some_baseline_model. The results are in .txt format and named by the AUROC score on the validation dataset and the corresponding hyperparameters of baseline models.

2. Then, you can run the INAFEN with a specific parameter setting with python main_INAFEN.py --dataset_name dataset_31_credit-g.csv --DTFT True --ARFC True --BMKD True --FC_min_support 0.2 --FC_min_confidence 0.8. The teacher model can be chosen by modifying the code teacher_model = XGBClassifier(n_estimators=100, max_depth=3, gamma=1) in main_INAFEN.py with the best teacher model according some metric on the validation dataset.

   The generated results can be found in ./Results/dataset_name/some_setting_of_INAFEN, in which some_setting_of_INAFEN can be DTFT[True]-ARFC[True]-BMKD[True], for example.