TBP-XFE: A transformer-based explainable framework for EEG music genre classification with hemispheric and directed lobish analysis

• A novel EEG music genre dataset with five classes was curated from 2,912 signals. • The proposed TBP-XFE framework achieved 93.2%–100% accuracy using 10-fold CV. • TBP combines transformer-based and transition table-based feature extraction for robust EEG representation. • DLob and hemispheric symbolic languages provided interpretable neural insights for music perception. Electroencephalography (EEG) signals offer important information for machine learning. In this work, we evaluate whether EEG signals can be used to classify music genres. We use a new transformer-based feature extraction method called the Three-Body Pattern (TBP). We also collected an EEG based music dataset containing five classes and these classes are: classical, popular, rap, ballad, and resting. Our feature engineering framework operates in four phases. First, the TBP method transforms each EEG signal to extract distinct features. Second, we use cumulative weighted neighborhood component analysis (CWNCA) to select the best features. Third, a t-algorithm-based k-nearest neighbors (tkNN) classifier assigns class labels. Finally, we apply Directed Lobish (DLob) and hemispheric symbolic languages to produce clear and explainable results. The TBP-related explainable feature engineering (XFE) framework achieved over 90% classification accuracy on the EEG music dataset. This represents a promising advancement in EEG based music classification because it produces clear and explainable outputs.