Construction of a Music Genre Preference Recognition Model Based on Deep Learning

Area of Science:

• Neuroscience
• Computer Science
• Music Psychology

Background:

• Music therapy is a growing non-pharmacological intervention for mental health.
• Deep learning shows potential in predicting music preferences and recognizing emotions.

Purpose of the Study:

• To construct a deep neural network model (CNN+RNN/EEGNet) for music type preference identification.
• To evaluate the impact of user familiarity on prediction accuracy using EEG signals.
• To explore applications in personalized music recommendations and music therapy.

Main Methods:

• Collected electroencephalogram (EEG) data using a four-channel Muse S wearable device.
• Categorized music into rock, ballad, and folk, collecting EEG data and familiarity ratings.
• Trained and tested CNN+RNN and EEGNet models with 10-fold cross-validation.

Main Results:

• EEG data alone predicted music types with 82.28% accuracy.
• Incorporating user familiarity and a multi-level rating system improved overall accuracy to 94.94%.
• Individual music type prediction accuracy reached up to 99.15% with familiarity features.

Conclusions:

• Combining user familiarity with EEG data significantly enhances music preference prediction.
• The developed deep neural network model (CNN+RNN/EEGNet) is effective for personalized music recommendations.
• This approach holds promise for advancing music therapy applications.