Explicabilidade de Graph Neural Networks: Uma Análise Comparativa com SHAP

Abstract

Graph Neural Networks (GNNs) are effective models for learning from structured data such as graphs and are applied in diverse areas. However, their interpretability remains a challenge, especially in critical applications. Methods like SHAP (\textit{Shapley Additive Explanations}) provide insights into predictions, but their application to GNNs is complex due to the relational structure of the data. This work reviews explanation approaches for GNNs, with an emphasis on those aligned with SHAP. The analysis shows that the methods, although similar in concept, differ in the choice of graph component used for the explanation among nodes, edges, or subgraphs and in the approximation techniques for calculating Shapley values. It is concluded that the choice of the appropriate method resides in the nature of the chosen task and the type of understanding sought.