Využití metod vysvětlitelného strojového učení při klasifikaci textů

This master thesis focuses on the research and application of explainable machine learning methods in the field of text classification. Special attention is paid to the implementation and subsequent comparison of the LIME method with the newly developed SMER method. Theoretical part of the thesis is focused on the research of current methods of explainability of machine learning models, methods of text preprocessing and representation. In the practical part, the LIME method is implemented in combination with a trained classifier and the Word2Vec model, which allows for obtaining local and then derived global explanations. These results are then compared with the SMER method using the AOPCglobal evaluation technique. During the comparison, correlations, and variations between the two methods are examined and differences between local and global importances are analysed. It was observed that in the analysed dataset, the local LIME importance was mostly not substantially different from the global importance for that word. Therefore, this thesis proposes a new visualization technique for communicating word importance, which utilizes global importance by default and reverts to local importance only when it is substantially different. The results also show that both LIME and SMER methods could have the ability to identify key features, with SMER appearing to be more accurate compared to LIME. The main contribution of this work is the integration of the results of both methods, to provide a more comprehensive view of word importance in text, with the possibility of clearly visualizing them. The paper concludes with a discussion of the obtained results, the identified limitations, and opportunities for future research.