Masterarbeit - Machine Learning: Concept Extraction Validation Benchmark

Field of study: computer science, mathematics, software design, software engineering, technical computer science or comparable.

Machine Learning (ML) models are reaching a maturity level that allows their operational use in businesses. However, in some areas, this use is limited by their ”black box” nature: the decision-making logic and potential errors of a model are not transparent, making it unsuitable for safety-critical applications or those requiring trust in the model. The field of Explainable Artificial Intelligence (XAI) addresses this by providing methods to make model behavior more interpretable. Among these, concept-based and prototype-based methods show promise in offering intuitive insights into model decisions. To truly build trust and ensure safe deployment of models, however, it is not enough for XAI methods to be intuitive — they must must also meet some key requirements. For example, the methods need to be reliable and their explanations need to be faithful to the model, while having a complexity level appropriate for human users. To ensure that these properties are met, XAI methods must be rigorously validated. Furthermore, such an evaluation should be systematic, allowing to compare most methods on the same ground. A framework for this is still largely missing in current XAI pipelines.

This thesis investigates the systematic benchmarking of concept-based explanation methods for machine learning models. It adapts an existing benchmarking framework, originally developed for pro- totype methods, to support the evaluation of concept-based explanations. The project also includes the empirical testing of concept extraction methods, evaluating their effectiveness and reliability using diverse metrics and datasets. The work contributes toward standardizing the evaluation of XAI techniques to ensure that generated explanations are meaningful and faithful to the underlying model.