Explicabilidade do modelo de aprendizado de máquina apoiada pela explicabilidade de dados: uma abordagem baseada em proveniência

Abstract

Artificial Intelligence solutions, especially those related to Machine Learning (ML), have achieved remarkable levels of performance due to the continuous advancement in computational capacity, the abundant availability of data, and the evolution of learning methods, which have become increasingly complex and sophisticated. To enhance the interpretability of ML models, Explainable AI (XAI) has emerged. XAI is of fundamental importance in increasing confidence in ML predictions and has become crucial for interpretation, especially in predictive models in critical domains. To provide a better understanding of the data, data provenance offers an explanation of its origin and derivation. Some studies have already explored the use of data provenance information in various stages of the ML lifecycle to contribute to explainability. However, there is a gap in the relationship between data provenance and the model’s explainability provided by Explainable Artificial Intelligence (XAI) techniques. In order to address this issue, this study proposes Explain- able Machine Learning Model supported by Pre-processing Provenance (xMML-PPP), an approach to capture provenance data, especially during the pre-processing phase, and relate it to the results of explainability techniques. To achieve this, a relational data model has also been proposed, which serves as the foundation for our data explainability concept. The main objective is to enhance the explainability of these techniques by complementing them with information derived from the pre-processing phase. For the application of this approach, a tool called xMML-PPP Tool has been developed, where various cycle information, including that from the pre-processing phase, is captured and stored in xMML-PPP Prov, the repository used by the tool to store the captured data, where, through queries to the data stored in xMML-PPP Prov, the information is retrieved. The approach was evaluated through two case studies, in which two experiments with different configurations were conducted for each of the case studies. This enabled the analysis of the interpretability behavior in different scenarios. The models were trained using the xMML-PPP Tool with the Random Forest algorithm, and the SHAP interpretability method was applied. The results of the experiments presented that the improvement in the explainability of ML models was mainly achieved through understanding the derivation of the attributes that constituted the model, enriched by data explainability.