A Novel Approach to xAI Techniques: Improving Accuracy and Interpretability of Tree-Based Models

The inherent processes that allow machine learning (ML) models to achieve their predictions are usually opaque to humans, specially to those who are not directly involved in the execution of those models. Without a trace of a doubt, ML models are going to be part of decision makers'  basic toolkit, if they are not yet. Explainable artificial intelligence (xAI) techniques are intended to assist the communication of the ML results to human decision makers. We suggest a novel approach within xAI techniques that leverages a Kohonen network of Shapley values, to offer valuable perspectives on tree-based models to the end-user. This approach can be used in several decision making frameworks, but we focus on some actuarial risk management applications related to savings products' policyholder behavior, such as the improvement of paid-up risk management.

The core of this presentation is based on the paper Bermúdez, Anaya, and Belles-Sampera (2023) [https://doi.org/10.1016/j.frl.2023.104242]