Erklärbare Künstliche Intelligenz - Steigerung der Nachvollziehbarkeit überwachter maschineller Lernverfahren

Artificial Intelligence (AI), and in particular the field of machine learning, is becoming increasingly important in our everyday lives. In the future, these methods will increasingly be used, for example, in medicine to help diagnose illnesses or in banking to detect money laundering. Reservations about the use of these methods are very often associated with their high complexity and the associated lack of explainability. Models generated by machine learning methods are considered a "black box" because they are usually incomprehensible to the user. There is a lack of insight into how the model generates the results. The research field of Explainable Artificial Intelligence (XAI) tries to design solutions that make entire models or specific model results comprehensible. This thesis contributes to the research field of XAI, especially in the subfield of explainable supervised machine learning. The first major contribution of the work is the design of a procedure model which defines different types of explanations that can be generated in this subfield. In total, five types of explanations are distinguished, and these can be roughly divided into model and instance explanations. Another major contribution of the work is its procedures for generating explanations. Based on this procedure model, two procedures were designed: a global surrogate model and a local surrogate model. Starting from a neural network, a global surrogate model is generated in the form of a comprehensible surrogate model. The surrogate is generated using regularization so that it satisfies both the properties of explainability and accuracy. The local surrogate model generates the explainability of a single instance starting from a black box. The goal is to generate an explanation that is close to the decision boundary of the original data instance. This type of explanation allows the user to generate preferred outcomes. Explanations that make models or specific model results comprehensible are an important step in the process of explainable machine learning. To investigate the comprehensibility of explanations, the utility of generated explanations must be investigated with users. Therefore, another major contribution of the thesis is the investigation of the comprehensibility of explanations in the context of user studies. Different tasks and user groups were included in these investigations.