Interpretace black-box prediktivních modelů zaměřená na kategoriální proměnné

The thesis evaluates the usability of major methods for interpreting black-box predictive models on categorical variables to provide a comprehensive comparison. The main goal is methods summarisation, application on categorical datasets, usability evaluation and comparison. The first theoretical part introduces categorical variables and their challenges in machine learning (ML). Afterwards, the concept of explainable AI (XAI) and its relation to the interpretability of black-box ML models are discussed. Interpretability is further described to provide readers with the necessary background knowledge to comprehend the context in the practical part. Subsequently, major methods for ML interpretability are presented, and the theoretical part concludes by defining observation criteria for methods usability on categorical variables. The following practical part applies interpretability methods on two black-box predictive models, each trained on a different categorical dataset. The methods are implemented in the interactive Jupyter notebook environment using Python. The results of the methods are initially interpreted in detail and then summarised for the usability evaluation. Each method is assessed according to predefined criteria highlighting its key features for categorical variables. Furthermore, the strengths and weaknesses of the methods in terms of categorical variables are also analysed. The conclusion of the thesis summarises the key findings of the research and discusses the goal fulfilment, thesis limitations or further research. The outcome is the usability comparison of interpretability methods on categorical variables, including evaluating their strengths and weaknesses for these variables.