Kredi Riski Modellerinde Açıklama Güvenilirliğinin Nicel Ölçümü: Shap Tabanlı Yeni Bir Stabilite Metrik Yaklaşımı

TÜBİTAK Projesi, 1002 - Hızlı Destek Programı, 2026 - 2027

Proje Türü: TÜBİTAK Projesi
Destek Programı: 1002 - Hızlı Destek Programı
Başlama Tarihi: Eylül 2026
Bitiş Tarihi: Eylül 2027

Proje Özeti

Credit risk modeling is critically important for the sustainability of financial systems, risk management, and the reliability of decision support mechanisms. In recent years, the high predictive performance of machine learning-based methods has led to their widespread use in credit scoring processes. However, the nonlinear and complex structures of these models reduce the transparency of decision-making mechanisms and make the need for explainability even more crucial. In this context, the SHAP (Shapley Additive Explanations) method has become one of the most widely used explainable artificial intelligence techniques in the literature for interpreting model outputs. However, it is observed that in the current literature, SHAP-based explanations are mostly accepted as directly reliable, but the extent to which these explanations change under different model types, data variations, and sampling strategies is not systematically examined. This situation creates a significant methodological gap because explainability is only considered through interpretability, and the reliability of explanations is not considered as an independent research problem. Especially in financial decision systems, the variability of explanations obtained under different conditions on the same data leads to uncertainty about which factors the model relies on for its decisions, and thus weakens the reliability of decisions. This project aims to develop a novel stability metric (SHAP Stability Score – SSS) that can quantitatively measure the reliability and consistency of SHAP-based explanations, thereby filling the aforementioned gap. The proposed metric defines explanation stability through three complementary components: order consistency, explanation size variance, and sign consistency, and combines these components under a normalized mathematical function to produce a single integrated score. This approach, for the first time, brings together stability measures, previously considered separately in the literature, under an integrated structure, allowing for a multidimensional and comparable evaluation of explanation reliability. The project will utilize algorithms representing different model complexity levels, such as Logistic Regression, Random Forest, and XGBoost/LightGBM; systematic data variations will be generated through data splitting strategies, bootstrap sampling, controlled noise addition, and feature perturbations. Within this experimental framework, SHAP explanations produced by the same model under different data conditions and by different models on the same data will be analyzed comparatively, and the results will be evaluated using the proposed SSS metric. The developed metric will not be limited to credit risk modeling but has a wide range of applications, including evaluating the reliability of model descriptions in healthcare, insurance, marketing analytics, fraud detection, and other data-driven decision support systems. In this respect, the project contributes to making explainable AI systems transparent, auditable, and reliable, addressing a significant need, particularly in areas with high regulatory requirements. Furthermore, this study offers a strong interdisciplinary research framework, positioned at the intersection of mathematical modeling, machine learning, and finance. The developed FAQ metric will not only contribute to the academic literature but also strengthen trust in AI systems by increasing decision reliability in real-world applications. In this way, the project aims to lay the groundwork for the development of a new reliability-focused research paradigm in the field of explainable AI.