Author: Alexey Firstov, Data Scientist, Machine Learning Laboratory of Alfa-Bank. Presentation at.
The presentation explains where the bias in the training sample comes from in the credit scoring task and how the Reject Inference approach reduces it: internal loan approvals are supplemented with a "challenger" sample and an external loans sample, and the neural network is trained on the combined samples with tuned sample weights. According to the author, the result for the bank is +2 Gini points on the card model, which is almost 600 million (currency not specified) additional profit.
Continue reading — it's free
Google or a one-time email code. No password.