fuzzy-augmentation-reject-inference

Fuzzy augmentation for reject inference is a through-the-door data augmentation technique that duplicates rejected applicants as both good and bad outcomes, then assigns probability-of-default-based sample weights to build a complete training set.

To build a through-the-door training set, score rejected applicants using a fitted PD model with predict_proba, duplicate each record as both positive and negative outcomes, assign PD-based sample weights, and combine them with accepted data for training.

Yes, you can integrate the generated through-the-door dataset with autogluon-tabularpredictor-fit to train downstream predictors on the augmented data using the assigned sample weights.

Yes, you need a fitted ri_model equipped with predict_proba to estimate probability-of-default scores for rejected applicants, which are then used to calculate the sample weights for the dual-record augmentations.

The through-the-door data augmentation process requires the pandas library for data manipulation, along with the provided scripts to generate the augmented dataset and summarize the results.

Use fuzzy augmentation when you need to preserve distributional uncertainty by weighting rejected applicants as both good and bad outcomes simultaneously, rather than forcing a single binary classification for each rejected record.