tabular_augmentation contains some classical and novel methods used for data augmentation, making tabular data
augmentation easier, especially for few-shot learning case.
SMOTE-based methods
from tabular_augmentation import smote_augmentation
method = 'SVMSMOTE'
x_synthesis, y_synthesis = smote_augmentation(x_few_train, y_few_train, method, seed=seed,
oversample_num=100, positive_ratio=None,
knn_neighbors=3)
tabular_model_test(x_synthesis, y_synthesis, x_test, y_test, model_name='xgb')Mixup-base methods
from tabular_augmentation import mixup_augmentation_with_weight
method = 'vanilla'
x_synthesis, y_synthesis, sample_weight = mixup_augmentation_with_weight(
x_few_train, y_few_train, oversample_num=200, alpha=1, beta=1, mixup_type=method, seed=seed, rebalanced_ita=1)
tabular_model_test(x_synthesis, y_synthesis, x_test, y_test, model_name='xgb', sample_weight=sample_weight)CTGAN/TVAE-based methods
Methods(CTGAN/TVAE/DeltaTVAE/DiffTVAE) use sdv_synthesis function to generate synthetic data, and ConditionalTVAE use sdv_synthesis_cvae function
from tabular_augmentation import sdv_synthesis, sdv_synthesis_cvae
method = 'CTGAN'
x_synthesis, y_synthesis = sdv_synthesis(
x_few_train, y_few_train, method, oversample_num=5000,
seed=seed, init_synthesizer=True, positive_ratio=0.5,
)
tabular_model_test(x_synthesis, y_synthesis, x_test, y_test, model_name='xgb')TabDDPM-based methods
from tabular_augmentation import ddpm_synthesis
method = "DDPM"
x_synthesis, y_synthesis = ddpm_synthesis(
x_few_train, y_few_train, method, oversample_num=5000, seed=seed, init_synthesizer=True, positive_ratio=None, train_steps=10000)
tabular_model_test(x_synthesis, y_synthesis, x_test, y_test, model_name='xgb')For details, please refer to example.ipynb
[ICLR' 18]mixup: BEYOND EMPIRICAL RISK MINIMIZATION Mixup
[ICLR' 22]Noisy Feature Mixup NoisyMixup
[ECCV' 20]Remix: Rebalanced Mixup
[NIPS' 19]Modeling Tabular data using Conditional GAN CTGAN
[ICML' 23] TabDDPM: Modelling Tabular Data with Diffusion Models TabDDPM