Audio feature-based user profiles for personalized music recommendation: A dataset-driven evaluation

In this paper, we propose an experimental content-based track recommendation system, which relies on an aggregated features which considers audio feature values from Spotify Data Catalog, track lyrics and popularity ratings. As system evaluation protocol, we evaluate how relevant the top-5 recommended tracks are to the user profile, which is based on average audio feature (danceability, energy, valence, loudness, instrumentalness, liveness, speechiness and acousticness) values from the user?s listening history. Based on the evaluation results, we come to the conclusion that the recommended tracks are similar to the user?s profile.