Gnasso, Agostino (2025) E2Tree: A methodological proposal to explain tree-based ensemble decision rules. [Tesi di dottorato]

Documento PDF Tesi_Gnasso.pdf Download (4MB)

Abstract

The aim of this work is to provide a methodological contribution on ensemble methods. Ensemble learning methods constitute a cornerstone of modern supervised learning, providing state-of-the-art predictive performance in both classification and regression tasks. Among them, Random Forests have emerged as one of the most effective algorithms, combining robustness, flexibility, and high accuracy. Yet, the very mechanism that grants Random Forests its predictive power (aggregating the output of numerous decision trees) renders the model largely opaque. This opacity hampers interpretability and explainability, turning Random Forest into a black box model and limiting its applicability in domains where transparency is essential, such as finance, medicine, and the social sciences. This dissertation presents a novel methodology, Explainable Ensemble Trees (E2Tree), which is a significant advancement for the field of explainable artificial intelligence (XAI). E2Tree merges the intuitive structure of decision trees with the predictive strength of Random Forests. By exploiting a dissimilarity matrix derived from co-occurrence patterns of observations within terminal nodes, the proposed methodology reconstructs a unified tree-like structure that effectively summarizes the ensemble process. In doing so, E2Tree preserves the predictive accuracy of the Random Forest model while providing a transparent, explainable, interpretable, and graphical representation of the relationships between predictors and outcomes. The dissertation advances this methodology through five main directions. First, it introduces the fundamental concepts necessary for comprehending the methodological and theoretical framework underlying the proposal. Second, a critical evaluation of existing interpretative strategies for Random Forests is presented, with the shortcomings of these strategies in capturing both local and global model behaviour being identified. Third, it formalizes the E2Tree framework in classification settings. Fourth, it demonstrates the explanatory capacity of the method through a benchmark application in the classification domain. Finally, the dissertation extends the framework to regression tasks by introducing novel node-level measures of goodness of fit and validating the approach through appropriate statistical tests, including the Mantel test. By offering a unified and explainable framework for ensemble tree models, E2Tree has the potential to bridge the long-standing gap between predictive performance and explainability. Its intuitive structure can enhance decision-making in domains where accountability and transparency are critical, foster greater trust in machine learning applications, and serve as a versatile tool for researchers to understand complex interactions within data-driven analyses.

Downloads

Actions (login required)

Modifica documento