Investigação da programação genética para explicabilidade em modelos de aprendizado de máquina

• Ano de defesa: 2024
• Autor(a) principal: Leonardo Augusto Ferreira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais
• Palavras-chave em Português: Engenharia elétrica; Aprendizado do computador; Inteligência artificial; Programação genética (Computação)
• Link de acesso: https://hdl.handle.net/1843/78202

Advances in Machine Learning (ML) are transforming how researchers conduct science in sensitive domains such as healthcare, education, justice, and criminal investigation. In response to this transformation, Explainable Artificial Intelligence (XAI) has emerged as a crucial research topic. This paper presents an innovative method called GPX (Genetic Programming Explainer), based on Genetic Programming for symbolic regression, aiming to provide clear and local explanations for decisions made by AI systems. GPX generates a set of samples in the neighborhood of the prediction to be explained and creates a local explanation model. The tree structure generated by GPX provides a symbolically, analytically comprehensible, and possibly non-linear expression that reflects the local behavior of the complex model. The use of partial derivatives from Genetic Programming results allows GPX to effectively communicate feature importance, producing user-friendly explanations. Additionally, the method can formulate counterfactual explanations, offering deeper insights into model behavior. Through comprehensive experiments on diverse datasets, GPX demonstrated excellence in four crucial aspects of XAI: providing understandable arguments, maintaining fidelity in classification and regression tasks, ensuring explanation stability, and promoting novel explanations. Compared to existing techniques such as LIME, GPX rivals or surpasses these approaches, as demonstrated by fidelity metrics. An innovative aspect of this work is the introduction of cosine similarity as an XAI strategy, which increases trust in the provided explanations. Despite the stochastic nature of Genetic Programming, the explanations generated by GPX remain stable, as validated by the stability metric. In summary, GPX shows promise in delivering informative arguments, ensuring fidelity and stability of explanations, and fostering the generation of novel explanations. This work represents an advance in XAI methods and encourages the exploration of symbolic regression via Genetic Programming to enhance explainability.