Investigating the Quality of Explainable Artificial Intelligence: A Survey on Various Techniques of Post hoc

As the time is passing, our dependency on intelligent machines is continually growing, which demands for more interpretable and transparent models. So, in any specific department, the real standard of artificial intelligence is judged by only if artificial intelligence is capable of interpret the model’s working which will generate users’ trust. Basically, explainable artificial intelligence or XAI targets to give a proper explanation of any particular machine-learning system, known as ML also, which empowers users (which are basically humans) to think, completely trust, and generate explainable models as much as possible. Choosing a suitable method for creating an XAI-enabled application needs a proper and deep perceiving of the basic logics within XAI and the associated methods. Among all the methods, black box-based artificial intelligence or AI method, for example, DNN or deep neural networks, has been broadly used for constituting prototype which are predictive and can interpretate typical relationship inside a dataset and can be used for making predictions for new unidentified data items. By using post hoc methods, one can explain inner working of these complex decision logic which is hidden from user. Basically, methods which are based on post hoc, approximate the working of black box nature by fetching rapports between predictions and values having different features. In this article, we discuss different XAI methods, specially post hoc method on different data item set. Using this taxonomy, also explain the various scenario on which post hoc can be applied and also elaborates how Post hoc provides better understandability and interpretability to users. This taxonomy can be used as a reference and comprehensive review of XAI technique qualities and elements for novices, researchers, and practitioners. As a result, it offers the framework for future research that is focused, use-case-oriented, and sensitive to context.