A Deep-Learning Approach to Emotion- Based Music Recommendation using Nigerian Facial Expression Data

Music has long been recognized as a powerful medium for evoking and regulating emotions. In recent years, the rapid growth of music streaming services and advancements in affective computing have opened new avenues for emotion-aware music recommendation systems. This study presents an emotion-based music recommendation system that analyses users’ facial expressions to suggest songs matching their emotions. 2, 073 faces image data were collected and hand-labeled (1, 063 “Happy,” 1, 010 “Sad”) captured in Nigeria and 36, 406 song records (happy vs. sad) via the Spotify API. A transfer-learning classifier based on VGG16 with frozen convolutional layers and fine-tuned dense layers was trained for 30 epochs. The model achieved 75.5% training accuracy and 73.7% validation accuracy. (recall: Sad 79%, Happy 72%), f1-scores of 75% and 76% for the happy and sad classes, respectively. FER engine was integrated into a Flask web interface that streams the user’s face in real time, infers emotions, and queries the music database to play matching tracks. Evaluation metrics revealed moderate class balance, with an overall accuracy of 75%, precision ranging from 73 - 78%, and The findings underscore the potential of integrating FER into music recommendation systems, enabling more intuitive and emotionally intelligent user experiences.