Model-Agnostic Interpretation via Feature Perturbation Visualization

• Wilson E. Marcílio Júnior UNESP
• Danilo Medeiros Eler UNESP
• Fabrício Breve UNESP

Resumo

As machine learning algorithms increasingly replace traditional approaches, ensuring their reliability becomes crucial in applications where incorrect decisions can lead to serious consequences. This work proposes a novel model-agnostic in-terpretation approach using feature perturbations, along with a validated visualization tool. The approach enables better understanding of model decisions by domain experts, facilitating effective decision-making in real-world applications.