A novel approach using explainable prediction of default risk in peer-to-peer lending based on machine learning models

Abstract

Online peer-to-peer (P2P) lending has expanded substantially during the previous decade globally. However, this quick expansion poses several potential risks as loan default risk in P2P lending remains unavoidable. As P2P lending has grown in both size and complexity, the challenges have also multiplied, leading to several complications, including high number of features, low-performing classification models and imbalanced dataset. Furthermore, machine learning models encounter another challenging issue known as the black-box problem. To overcome these challenges, the present work introduces a novel approach that involves tackling the dataset balancing issue using synthetic minority oversampling technique (SMOTE), employing carefully selected feature selection approaches (maximum relevance minimum redundancy (MRMR), sequential forward selection (SFS) and adaptive boosting (AdaBoost)) and machine learning such as nonlinear model (K-nearest neighbour (KNN)), tree-based model (random forest (RF)) and deep learning (multi-layer perceptron (MLP)). Compared to the previous studies, the present results showed that RF exhibited outstanding performance of 0.94, 0.94 and 0.99 in accuracy, F1-score and AUC, respectively. To address the black-box issue of the prediction model, enhance its interpretability and boost user trust, local interpretable model-agnostic explanations (LIME) and Shapley additive explanations (SHAP) explainable machine learning models were applied to the RF prediction model to elucidate its results. Furthermore, LIME and SHAP explainable machine learning models were applied to the RF prediction model, both with and without SMOTE resampling, to examine the influence of SMOTE resampling on the interpretability analysis of the RF prediction outcomes.