Classifying demonstration format and presenter identity in imitative learning task: EEG-based explainable machine...

This study investigates whether EEG signals can distinguish between different formats of gesture demonstration (live vs. video) and between individual demonstrators (one male, one female) in imitation learning tasks, using explainable machine learning approaches. EEG was recorded from 83 participants during three task types: observation, execution, and simultaneous observation-execution. Relative power in the alpha and beta bands was extracted from 31 electrodes. Classification was performed using Random Forest (RF) and Multilayer Perceptron (MLP) models, with hyperparameter optimization via Bayesian methods. SHAP (SHapley Additive exPlanations) values were used to interpret the contribution of individual features. Results showed that beta-band activity provided the most informative input for classification. MLP models demonstrated higher accuracy in identifying the demonstration format across tasks, reaching up to 77 % in the observation condition. RF models were more effective in distinguishing between individual demonstrators. SHAP analysis revealed that live demonstrations elicited more localized mu desynchronization in motor regions and increased beta activity in temporal and posterior temporal areas, consistent with stronger engagement of social-perceptual networks. In contrast, video demonstrations were associated with higher beta activity in prefrontal regions, suggesting increased reliance on cognitive processing. Notably, MLP models better captured features linked to video demonstrations, while RF models were more sensitive to features distinguishing live interactions and individual presenters. However, findings related to gender classification should be interpreted cautiously, as they are confounded by individual identity.