A New XAI Evaluation Metric for Classification
Asmaa M El-gezawy, Hatem M. Abdelkader, Asmaa Ali · IJCI International Journal of Computers and Information · 2023
Explainable AI (XAI) has become a hot topic across multiple sectors. In practical applications, classification models are severely constrained by the absence of transparency, which undermines trust and has a black-box nature, leading to a range of problems. Classification models necessitate the use of XAI approaches to address these limitations effectively. The Mean Evaluation of Metrics Change (MEMC) is a novel metric introduced in this research for evaluating the performance of Explainable AI techniques on a global scale, like post-hoc and intrinsic XAI for classification techniques on tabular data. The proposed MEMC metric is formed from a combination of the existing standard evaluation measures used for evaluating classification. The proposed MEMC has proven to be the convenient metric for determining the best explainer for a produced classification. The proposed MEMC metric is validated using a heart dataset from the healthcare sector. The experimental results show that the Artificial Neural Network (ANN) approach performed effectively on the heart dataset as an intrinsic XAI in machine learning. Deep Neural Network (DNN) also performs better as an intrinsic XAI technique when applied to this dataset. Furthermore, ANCHORS has shown strong performance as a post-hoc XAI technique when Random Forest (RF) and XG-Boost are used as classification models.