A Comprehensive Survey of Explainable Artificial Intelligence in Machine Learning Models

Abstract

Scientists with diverse interests, ranging from scientists, the European Union, and researchers focused on artificial intelligence, are reemphasizing the importance of interpretation capabilities of models in the face of the increasing proliferation and social impact of AI systems. The reason is that the AI systems might be acting overly deterministic and overly confident, sometimes misleading in areas of critical expertise. The interpretability is the capability of humans to understand the results and decisions of models. It is particularly concerned when the computational models make high-stakes decisions (for instance, health care) that favor uncertainty promotion or when matter a further discussion with different stakeholders about the modus operandi. For example, overly confident models may make erroneous predictions that result in inaccurate diagnosis results in health care applications. Or, in fraud detection, they may overlook the profile of the real fraudulent behavior, excluding determining indications that might lead to identifying the external sectors in the highest risk.