eXplainable artificial intelligence (XAI) aims to provide a reasonable explanation of the decisions taken by black-box AI models. This ensures the transparency of the employed models by transforming them from black-box to white-box models, which can be safely employed in practice. Regarding the XAI taxonomy, the XAI methods can be classified into:
- Model-agnostic vs model-specific: Model-agnostic XAI schemes are independent of the internal architecture of the black-box model including the weights and the hidden layers. Whereas model-specific schemes depend on a specific model like feedforward neural network (FNN) or convolutional neural network (CNN) and cannot be generalized to any other model. Therefore, model-agnostic schemes are characterized by their high flexibility and can be used despite the type of the considered model.
- Local vs global: Local XAI schemes are those that generate explanations for a group of samples; thus, they are highly dependent on the utilized dataset. In contrast, global XAI schemes generate explanations that are related more to the model behavior.
- Pre-model, in-model, and post-model strategies: XAI schemes can be applied throughout the entire development pipeline. Interpretability can therefore be acquired in three main phases. Pre-modeling explainability is used to define the useful features of the dataset for a better representation. Hence, pre-modeling aims to perform exploratory data analysis, explainable feature engineering, and dataset description. In contrast, in-model explainability is to develop inherently explainable models instead of generating black box models. Finally, the post-model explainability method extracts explanations that are dependent on the model predictions.
In future publications, we will elaborate on different XAI categories, providing detailed examples and highlighting the main advantages and disadvantages of each category