• topology learning
• EEG
• MAHNOB
• affective computing
• concept-wise
• DEAP
• emotion recognition
• XAI
• concept selection
• artificial neural networks
• machine learning

Providing an understandable explanation of the model result is particularly important in the medical field. In contrast to the concept of AI algorithm as a "black box", XAI models expose the internal logic behind the outcome of AI algorithms, making this process understandable by humans as required by EU regulations. This thesis focuses on the design and testing of an XAI approach aimed at analyzing the subject's EEG, with the aim of classifying emotions in an explainable way
The proposed solution involves a concept-wise two-level architecture, where in the first level, a neural network tries to represent the EEG data of the subjects in an emotional space using concept-wise classifiers, where the concepts are the emotions. These features in the emotional space are obtained using similarity learning techniques.
While in the second level of architecture, a predictive model of machine learning deals with solving a classification problem starting from the emotional features provided by the first level. Here, several classifiers, one for each class, in a one-vs-all configuration, undergo a learning phase that is performed using a concept selection procedure, where all the possible combinations of the concepts of the emotional space, created in the first level, are passed to these neural networks. In this way, the topology of the architecture will be the explanation of the phenomenon considered, as each neural network will learn which concepts are the best for solving the classification. In other words, the model itself tells us which are the most important concepts that are used to identify the emotions or more generally the classes of the considered classification problem. In this way we can understand part of the logic behind the classification process.
We focus on the prediction of emotions by considering problems such as the arousal - valence classification at different levels. Arousal and valence are two metrics used to describe an emotional state, in terms of intensity and positivity/negativity of the emotion itself.
The proposed solutions beat those present in literature, in particular in the case of the arousal and valence two-level (low and high) classifications, reaching an accuracy of about 87% and 91%.
Mainly we carried out subject-independent experiments, with an architecture capable of taking EEG data from different subjects as input, and we used both 10-fold cross-validation and leave-one-subject-out cross-validation. We used two publicly available datasets which are MAHNOB-HCI and DEAP.