B0816
Title: Unlocking explainable in ensemble trees
Authors: Massimo Aria - University of Naples Federico II (Italy)
Agostino Gnasso - University of Naples Federico II (Italy)
Carmela Iorio - University of Naples, Federico II (Italy) [presenting]
Giuseppe Pandolfo - University of Naples Federico II (Italy)
Abstract: Explainability is the capacity to provide understandable explanations to human beings regarding the processes occurring within a model, from input to output. Ensemble methods refer to supervised learning algorithms that leverage multiple models to yield highly accurate solutions. In the regression and classification problems, random forest (RF) stands out as the most commonly employed technique. RF represents an effective method of ensemble learning that offers a combination of accurate prediction and flexibility. Although it is widely regarded as an intuitive and transparent approach to model construction, it is also categorized as a black box model due to the numerous complex decision trees it generates. To address this issue, a theoretical framework known as ExplainableEnsemble Trees (E2Tree) is proposed. It presents two key advantages: (i) constructing an interpretable decision tree that ensures the predictive performance of the RF model and (ii) providing the decision-maker with an intuitive graphical structure for managing the model. The approach aims to visually represent the intricate relationships and interactions among the variables utilized in the model. By leveraging the strengths of decision trees and random forest models, a dendrogram-like structure is introduced that comprehensively explains the information encapsulated in the random forest output.
