Distributional Feature Importance for Correlated Predictors

Koca Güler B., Yozgatlıgil C.

22nd International Conference Applied Statistics 2026, Koper, Slovenya, 21 - 23 Eylül 2026, ss.70, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Koper
  • Basıldığı Ülke: Slovenya
  • Sayfa Sayıları: ss.70
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet


In the literature, there exists a rapidly increasing trend in the use of machine learning and deep learning models thanks to their high predictive power and their ability to handle non-linearity and complex structures. However, because of their high number of dimensions, the inner workings of the models are obscured, making the model decisions mysterious. Model explanation can take several forms such as rule-logic extraction and counterfactual explanations, however explainability through feature importance is the most widely adopted strategy. SHAP and impurity-based importance are examples of the primary eXplainable AI (XAI) methods targeting feature importance. However, they suffer from the high level of relation between the variables, which is called multicollinearity, and these methods distribute marginal contributions among correlated variables. Hence, in this study, we explore a framework that focuses on feature importance under multicollinearity while mainly drawing on optimal transport theory, specifically the Wasserstein distance together with structural information among the variables. Rather than relying on marginal effects alone, the framework quantifies the feature importance based on distributional shift in the model’s behavior. Preliminary findings suggest that this integration is a promising alternative for explainability and interpretability, especially in the presence of multicollinearity.