Green AI in the finance industry: Exploring the impact of feature engineering on the accuracy and computational time of Machine Learning models

Abstract

As research and practice in Artificial Intelligence (AI) applications rapidly expand, the support for AI deployment is also increasing. While the abundance of data allows for sophisticated feature engineering techniques that can enhance accuracy, it is crucial to highlight both the computational costs and the efficiency with which these models operate. This paper compares the processing time and accuracy of individual and hybrid Machine Learning (ML) models in predicting customer loyalty within financial contexts. Frameworks that incorporate feature engineering and green AI principles are used separately in both individual and hybrid approaches. The individual models are the commonly used regressor-based algorithms applied to business problems. The hybrid models first use k-Means to cluster customers, followed by the application of individual regressor-based models (e.g., decision trees, gradient boosting, and LightGBM). The present results show that using fewer features results in only a marginally lower accuracy compared to models with more features (a difference of ≈0.01 in MAE when comparing the use of 18 versus 85 features). Additionally, this article clearly demonstrate the trade-off between higher accuracy and longer computational time in hybrid ML models versus lower accuracy and shorter computational time in individual models when predicting customer loyalty. Hybrid models exhibit a lower MSE (≈0.88) compared to individual models (≈0.91). These findings provide managers with insights on selecting the most appropriate model based on their organization’s specific needs.