Aplicação do Algoritmo LIME para explicar como classificadores black-box usam atributos na tomada de decisão

Abstract

With the recent popularization of generative artificial intelligence, there has been an increase in concern about making understandable the determining criteria used by models to answer questions proposed by users. In this work, we sought to use Explainable AI (XAI) techniques to clarify the functioning of black-box models, those in which the path taken by the algorithm to find the answer offered is not clear. The concern about the explainability of an AI model is mainly to ensure reliability and monitor possible ethical and rights failures, such as: prejudice, discrimination, invasion of privacy, among other problems that training can cause. The black-box model chosen was the Multilayer Perceptron (MLP) and we sought to increase its explainability by applying the LIME Algorithm. The application aimed to understand the weight assigned to the attributes in the MLP decision. To this end, a weighted kNN algorithm was adjusted with the assignment of the weights found by the importance ranking of the attributes generated by LIME when applied to the MLP. Finally, performance measures such as execution time, precision, accuracy, recall, F1 Score and Jaccard were analyzed to compare the MLP with a white-box classifier, kNN. The results obtained suggest that there is an increase in the explainability of the MLP using XAI techniques and that the average performance of LIME improves the expressiveness of the global influence of the attributes in the classification model.