---

name: fuzzy-augmentation-reject-inference description: Build a Through-the-Door training set with reject inference using fuzzy augmentation, including PD-based sample weights; pairs with autogluon-tabularpredictor-fit for modeling the augmented data.

Fuzzy Augmentation Reject Inference

Purpose

Create a weighted Through-the-Door dataset by scoring rejected applicants with a KGB model, duplicating them as good and bad outcomes, and assigning sample_weight by predicted PD.

Usage

• "apply fuzzy augmentation"
• "build TTD dataset with reject inference"
• "weight rejected applicants by PD"

Instructions

1. Train a logistic regression model on accepted data to estimate PD.
2. Score rejected applicants to get PD values.
3. Create two copies of rejected rows:
   • Copy A: default_flag = 1, sample_weight = PD
   • Copy B: default_flag = 0, sample_weight = 1 - PD
4. Combine accepted data with both copies of rejected data and add a source column.
5. Use ./scripts/create_ttd_data.py to standardize the augmentation.
6. Summarize results using ./templates/ttd_summary.md.