Skewed fat-tailed distributed (imbalance or class-imbalance) datasets pose over- whelming aberrations in numerous machine learning (ML) algorithms, particularly in real-life applications, especially in the domain of credit risk modelling, where default cases (minority-classes) are often outnumbered by non-default cases (majority-classes) cases or vice versa. Data-level (DL) approaches have been suggested in the recent literature as remedies for skewed fat-tailed distributed datasets. The popularized DL approach in contemporary studies is the synthetic minority over-sampling technique (SMOTE) and its variants that are capable of mitigating the risk of overfitting and minimizing the generalization errors. However, these approaches can introduce noisy instances that adversely diminish the robustness of the ML algorithms. Also, they are often amenable to the presence of nominal features with mismatching labels that are inherent in real-world datasets. To bridge these gaps, we proposed a hybrid innovation framework that effectively mitigates the aberrations presented by nominal features with mismatching labels and noisy instances simultaneously. The proposed approach is the SMOTE-edited nearest neighbors-encoding nominal and continuous (SMOTEENN-ENC) features. The efficacy of our novelty was evaluated against DL approaches suggested in the literature, orchestrated to handle skewed fat-tailed distributed datasets with inherent diverse features. This approach was coupled with widely employed ensemble algorithms, namely the random forest (RF) and the extreme gradient boost (XGBoost). The results suggested that our novelty, SMOTEENN-ENC, integrated with the XGBoost algorithm demonstrated superiority and stability in the predictive performance when applied to skewed fat-tailed distributed datasets with inherent diverse features.
Skewed fat-tailed distributed (class-imbalanced) datasets with inherent heterogeneous features often induce overwhelming aberrations in numerous machine learning (ML) algorithm applications, particularly in credit risk assessments. To mitigate these complexities, data-level (DL) approaches have been suggested in the literature as remedies. Notably, the benchmarks of DL approaches in the literature consist of variants of the synthetic minority oversampling technique (SMOTE) customized to effectively handle various data aberrations using distance measures. To counter the limitations of the Euclidean distance (ED) in capturing heterogeneous features, their computational efficiency is enhanced through the utilization of the modified ED (MED). However, the MED substantially fails to account for correlations between features. To circumvent these shortcomings, some studies have suggested using the Mahalanobis distance (MD), which accurately accounts for correlations in features. Nevertheless, this measure considerably falls short due to its inability to capture heterogeneous features, as well as its susceptibility to outliers. Therefore, this study proposes a modified Mahalanobis distance (MMD) designed to adequately capture heterogeneous features. Additionally, the MMD parameters were estimated using the median absolute deviation (MAD), which efficiently discards outliers. The study evaluated the computational efficiency of SMOTE-based approaches combined with the MMD, computed intrinsically to MAD approach, using the most widely employed ensemble-based ML approach, the random forest (RF) algorithm. The findings indicate that our novelty significantly outper-formed conventional approaches. This investigation enriches the understanding of generalizable predictive performance in the credit risk landscape.