Minimum Covariance Distance-Based SMOTE Approaches for Zero-Inflated Datasets with High-Dimensional Heterogeneous Features in Big Data Analytics

Big data in the credit risk landscape may be characterized by zero-inflated datasets, heterogeneity and high dimensionality. These data aberrations adversely affect the predictive capabilities of conventional approaches, e.g., logistic regression (LR). Adaptations of the modified Euclidean Distance (MED)-based synthetic minority oversampling technique (SMOTE) approaches have been suggested in contemporary literature as remedies for zero-inflated datasets and heterogeneity. However, these approaches substantially fail to the capture variability and correlations among features, while remaining susceptible to outliers and collinearity, particularly in high dimensionality scenarios inherent in big data. To circumvent these complexities, we propose robust distance-based SMOTE approaches. We suggest extending the Mahalanobis Distance (MD) to the modified MD (MMD), computed intrinsically to the minimum covariance determinant (MCD) approach, for SMOTE approaches to achieve robustness in the big data phenomena. High dimensionality is often remedied by adaptive penalizations due to their ability to exhibit oracle properties. Therefore, this paper evaluates the efficacy of the MCD-based SMOTE approaches, leveraging MMD computed intrinsically to the MCD approach, in conjunction with adaptive penalized LR. The empirical evidence from this study suggests that this approach, demonstrates superior predictive performance. These findings confirm that our novelty significantly outperforms conventional approaches in big data scenarios. This paper’s originality stems from the interplay of MCD-based SMOTE approaches, utilizing the MMD computed intrinsically to the MCD approach with adaptive penalized LR. These contributions address data aberrations posed by zero-inflated datasets coupled with high-dimensional heterogeneity in modelling big data phenomena.