University of Twente Student Theses

Login

Explaining black box decision-making : adopting explainable artificial intelligence in credit risk prediction for P2P lending

Wijnands, M.P.J. (2021) Explaining black box decision-making : adopting explainable artificial intelligence in credit risk prediction for P2P lending.

[img] PDF
3MB
Abstract:P2P lending platforms are rapidly expanding and Artificial Intelligence (AI) is increasingly used to predict the the creditworthiness of borrowers, however these black box models lack explainability. We explore three state-of-the-art post-hoc model agnostic explainable AI (XAI) techniques named Local Interpretable Model Agnostic Explanations (LIME), Anchors, and Shapley Additive exPlanations (SHAP) and we come up with an exploratory framework called ARRGUS to assess the impact of the XAI techniques, as no assessment framework exists for it at present. We apply XAI to ML-based credit scoring models that are trained on classifying if a customer is creditworthy from the LendingClub dataset. We apply the Synthetic Minority Oversampling TEchnique (SMOTE) to deal with the class imbalance problem in the dataset. We present multiple comparisons for the different XAI techniques and discuss the results in detail. Our results show that all three XAI techniques provide fairly consistent explanations that are in line with financial logic. Based on ARRGUS, the SHAP technique scores best and is the most compliant. We conclude that XAI techniques generate explanations that are understandable for all users involved or affected by the outcome of ML models. Keywords: Explainable AI, Peer-to-Peer lending, Credit risk prediction, Machine Learning, LIME, Anchors, SHAP
Item Type:Essay (Master)
Clients:
Ernst & Young
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:31 mathematics, 50 technical science in general, 54 computer science, 70 social sciences in general
Programme:Industrial Engineering and Management MSc (60029)
Link to this item:http://purl.utwente.nl/essays/88621
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page