Biyomedikal metinlerin sözcük iliştirme ve sözcük torbası yöntemi kullanarak melez yaklaşımla sınıflandırılmasında açıklanabilir bir yapay zeka uygulaması

Since most of explainable AI system (XAI) evaluation measures in literature need human intervention and are subjective, the need for unbiased and more automatic methods is indispensable. This thesis proposes an XAI accompanied by a multi-class classification model using a hybrid feature representation method. We developed three quantitative XAI evaluation measures that objectively calibrate the complexity of two XAI methods, which are SHAP and LIME. In that context, we create a hybrid feature representation method that is used as a prepossessing step in a deep learning multi-classification system, combining both bag of words (BoW) and word embedding (WE) techniques. To this end, we use a bidirectional long-short-term memory (LSTM) deep learning model for the system. Thereafter, we implement several XAI methods, i.e. local and global XAI, on the outputs of the multi-classification system in order to explain the decisions of the employed model. Furthermore, we apply our proposed model and designed system to a medical multi-class benchmark called OHSUMED. Finally, we compare the complexity of the explainability between SHAP and LIME methods using the depth of the decision tree as an objective XAI evaluation measure. The results of the new quantitative and automatic measures proposed by us show that SHAP outperforms LIME method because it is less complex. Additionally, based on SHAP feature importance scores, we become able to identify which features are the most significant and hence must be included in the multi-class classification model. As a result, using both the hybrid feature representation system as a prepossessing step and our proposed XAI evaluation measures, the accuracy of the classification model increased from 45.4% to 92%.