Sunday 23 February 2025
Credit scoring models, used by banks and lenders to assess an individual’s likelihood of repaying a loan, have long been plagued by issues of bias and opacity. A new study aimed to tackle these problems by exploring the use of a dimensionality reduction technique called Linear Discriminant Analysis (LDA) in combination with machine learning algorithms.
The researchers analyzed a large dataset from Lending Club, one of the largest peer-to-peer lending platforms in the world, which contains information on over 2.5 million loans made between 2007 and 2020. They applied eight different machine learning models to the data, including popular techniques such as decision trees and neural networks.
The results showed that a hybrid model combining XGBoost, a powerful gradient boosting algorithm, with a deep neural network (DNN), achieved the highest accuracy in predicting loan status. However, this model was also found to be less transparent than others, making it difficult for users to understand why certain decisions were made.
To address this issue, the researchers used two different explainable AI techniques: Local Interpretable Model-agnostic Explanations (LIME) and Morris Sensitivity Analysis. LIME provides a local explanation of how individual predictions are made, while Morris Sensitivity Analysis ranks features by their influence on predictions.
The analysis revealed that certain features, such as the remaining outstanding principal and the number of delinquent accounts, were particularly influential in determining loan status. These findings provide valuable insights into the decision-making process of credit scoring models and could help to identify biases and errors.
The integration of LDA with machine learning algorithms was found to improve model performance while also reducing dimensionality and feature complexity. This is particularly important in large datasets like Lending Club’s, where dealing with a large number of features can be computationally expensive and lead to overfitting.
The study highlights the importance of balancing model performance with transparency and explainability in credit scoring applications. By using techniques like LDA and XAI, lenders can create more accurate and fair models that are also easier for users to understand.
The findings have significant implications for the financial industry, where credit scoring models are used to make critical decisions about loan approval and interest rates. As the use of machine learning algorithms becomes increasingly widespread, it is essential that these models be transparent and free from bias in order to maintain trust and confidence in the system.
Cite this article: “Improving Credit Scoring Models with Explainable AI Techniques”, The Science Archive, 2025.
Machine Learning, Credit Scoring, Linear Discriminant Analysis, Lending Club, Peer-To-Peer Lending, Transparency, Explainability, Xgboost, Deep Neural Network, Local Interpretable Model-Agnostic Explanations
