Interpretability in machine learning based on the Shapley value: Approximation methods and applications

• Grant number: 25/00700-0
• Support Opportunities: Regular Research Grants
• Start date: May 01, 2025
• End date: April 30, 2027
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Guilherme Dean Pelegrina
• Grantee: Guilherme Dean Pelegrina
• Host Institution: Escola de Engenharia (EE). Universidade Presbiteriana Mackenzie (UPM). Instituto Presbiteriano Mackenzie. São Paulo , SP, Brazil

Abstract

Frequently, practical situations require the application of complex machine learning methods to model the relation between input data and the system's output. Generally, these methods are difficult to interpret, which brings a disadvantage when there is a need to understand the impact of attributes on the obtained result. As a solution, recent works have been developing model-agnostic methods that can be used to extract interpretability from complex models. One example is the Shapley value-based models, a concept from cooperative game theory that provides the marginal contribution of each attribute to the performance of the trained machine. However, this approach has a difficulty. The computational complexity increases exponentially with the number of attributes. Therefore, the exact calculation of Shapley values becomes unfeasible in applications where the number of attributes is large. In this context, the first research focus of this project is the development of approximation methods for Shapley values. The idea is to explore both game theory concepts and data dimensionality reduction methods in order to reduce computational complexity. Once we achieve such a reduction, we will be able to use the Shapley values as a tool to extract interpretability in the signal processing problems addressed in the second research focus of this project. More specifically, we will investigate the importance of sensors used in the construction of medical signal acquisition devices, such as electrocardiograms, and the contribution of attributes extracted from audio signals for emotion recognition. (AU)