Implementation of Bidirectional Encoder Representations from Transformers in a Content-based Music Recommendation System for Digital Music Platform Users

Digital music platform users today have unlimited access to millions of songs from various genres and artists through music streaming services. However, with so many music platforms available, users often need help finding songs that suit their preferences. This study presents a music recommendation system that utilizes lyrical analysis to provide users with relevant song suggestions based on selected lyrics. The system employs a two-pronged approach: the Term Frequency-Inverse Document Frequency (TF-IDF) method for initial feature extraction and the IndoBERT model for advanced contextual representation of song lyrics. A dataset of 8,944 Indonesian language songs was compiled using scraping techniques from various sources. The recommendation process is driven by cosine similarity calculations between the lyrics of the selected songs and the entire dataset, enabling the identification of songs with similar themes and messages. Model evaluation through a five-fold Multi-Class Cross-Validation (MCCV) approach yielded promising results, indicating high precision, recall, and F1 scores. The study results show that the system built can provide recommendations with good precision performance with Precision@k values varying between 0.7965 to 0.8371, Recall@k values ranging from 0.8017 to 0.8204, and F1-score@k values varying between 0.8083 up to 0.8190. Overall, the model shows strength in providing accurate recommendations and good performance stability