Reproduce “Enhancing Sequential Music Recommendation with Personalized Popularity Awareness” on OpenResearch

About this paper

Davide Abbattista, Vito Walter Anelli, Tommaso Di Noia, Craig Macdonald, Aleksandr Vladimirovich Petrov

An approach is proposed to enhance sequential music recommendation by directly integrating personalized popularity information into Transformer-based models, addressing the repetitive nature of music consumption. This method yielded substantial performance improvements, ranging from 25.2% to 69.8% in NDCG, across multiple metrics and datasets.

Key insights

• Personalized Most Popular, a simple baseline based on user-specific historical listening frequency, often outperforms complex state-of-the-art Transformer-based models in music recommendation.
• Explicitly integrating personalized popularity information significantly enhances Transformer-based models, enabling them to capture the repetitive consumption patterns inherent in music.
• Models perform better when they learn to refine recommendations based on an existing personalized popularity distribution, rather than attempting to learn the entire distribution from scratch.

Summary by alphaXiv.