S-VAERec: A Hybrid Music Recommendation Strategy Based on Generative Modelling and Music Similarity

In the era of big data, music platforms require advanced algorithms to meet growing recommendation demands. This paper proposes a hybrid music recommendation framework, S-VAERec (Similarity-Variational Autoencoder Recommender), leveraging the intrinsic characteristics of music. The framework combines content-based and social-based similarity metrics to calculate item-user scores and employs a Variational Autoencoder (VAE) to extract latent features from historical data, generating additional predictive scores. These scores are linearly weighted to produce a combined rating prediction, enhancing recommendation accuracy. Experiments on the Last.fm dataset demonstrate fast model convergence and improved performance: S-VAERec reduces prediction error (RMSE) compared to similarity-based algorithms and increases Recall compared to generative model-based approaches. By integrating the strengths of both methods, S-VAERec achieves high accuracy and strong generalisation, offering a robust solution for personalised music recommendations.