J. Charles, S. Lekamge
Abstract
Music experience is closely associated with our moods and emotions. Even though data mining techniques have been widely adopted in computational analysis of music-emotion, traditional melodies such as Sri Lankan folk melodies are less explored computationally. Therefore, considering a Sri Lankan folk music dataset, we performed classification using Support Vector Machines, Naive Bayes, Random Forest (RF), k-Nearest Neighbor (k-NN), and Logistics Regression (LR), using a set of features pertaining to dynamics, rhythm, timbre, pitch, and tonality. k-NN yielded the highest accuracy (78.44%) while RF and LR yielded accuracies of 76.19% and 73.42% respectively. Combining the above three classifiers, an ensemble model was developed. Max-voting was applied and the results were further enhanced using ensemble boosting. With optimized features, AdaBoost (RF as base estimator) yielded the highest accuracy (95.23%) while reducing the training time significantly. Expanding the dataset in terms of number of music stimuli and emotion categories are looked forward.
Keywords: Music-emotion classification, Machine learning, Ensemble approach, Max-voting, Sri Lankan folk music
**View Full Paper in the Springer Source