Building a Benchmark Framework for eXplainable Artificial Intelligence (XAI) Methods

25/11/2022

Artificial intelligence (AI), namely its sub-fields machine learning and deep learning, have demonstrated impressive outcomes in a variety of scientific research domains, such as medicine, security, and finance. However, complex AI systems, despite demonstrating great results and accuracy performances, are seen as black-boxes that suffer from lack of explainability. Therefore, as AI systems continue to grow, it becomes important for humans to understand how each black-box arrived to a certain result. This way, the field of eXplainable artificial intelligence (XAI) arose from the necessity of solving the black-box problem. XAI field has been growing fast, but in different directions, revealing the difficulty the scientific community faces to agree on common definitions and evaluation criteria, which are often formulated in a subjective manner. To overcome this gap in research, the present dissertation proposes a benchmark framework for XAI methods, which is designed based on a methodological systematic literature review in order to derive objective and measurable performance indicators in a comprehensive and consensual manner. This framework is then applied to compare 9 well-known or promising XAI methods considering a tabular dataset from the medicine domain (heart disease prediction). This benchmark study showed the relevancy of the CIU method, which covers to a better extent the 10 selected properties of explainability, when compared to other methods. Moreover, the proposed framework contributes to the settlement of common formalism and taxonomy, which promotes the uniformity that is lacking in the XAI field.

Autores da comunidade :

• 

  D

  Dulce Marques de Carvalho Martins Canha

  ist189797

Orientadores desta instituição:

• 

  A

  Ana Luísa Nobre Fred

  ist12170

industrial-biotechnology - Biotecnologia Industrial

eng - Inglês

Instituto Superior Técnico