Título: Exploring Transformers in Embedding -Based Music Recommendation
Autores: Vinicius Sylvestre Simm, Douglas Rorie Tanno & Marcos Domingues
Resumo: Music recommendation systems help users explore large music collections by suggesting tracks they might enjoy. This work focuses on session-based music recommendations, where suggestions are made using only the sequence of songs played during a single listening session, without any prior user history. We compare three approaches for generating recommendations based on vector representations (embeddings) of songs. The first is a simple model that averages the embeddings of all tracks in a session. The second one uses an LSTM neural network. The third approach uses a Transformer neural network, which is better designed to understand the order and context of the songs. In all cases, the models generate a vector to compute the cosine similarity to all known item embeddings and to recommend the most similar tracks. These embeddings are created using the Word2Vec algorithm, with both Skip-Gram and CBOW variations. The results show that the Transformer model achieves better recommendation quality than the baseline, despite requiring more computational resources.
Palavras-chave: Recommendation Systems; Embeddings; Session-Based Recommendation Systems; Neural Networks; LSTM; Transformers.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1174261
Artigo em PDF: CBIC_2025_paper1174261.pdf
Arquivo BibTeX:
CBIC_2025_1174261.bib