Similarity of nearest-neighbor query results in deep latent spaces

Abstract

Music recommendation systems are commonly used for personalized recommendations. However, there are cases where due to privacy concerns or design decisions, there is no user information nor collaborative filtering data available. In those cases, it is possible to use content-based similarity spaces to retrieve the most similar tracks to be recommended based on the reference track. In this paper, we compare the latent spaces extracted from state-of-the-art autotagging models in terms of the similarity between lists of retrieved nearest neighbors. We additionally study item factors from collaborative-filtering data as a reference. We provide insights into how much the choice of the architecture, training dataset, or model layer (output vs. penultimate) as well as a projection of the latent space onto 2D changes the list of retrieved nearest neighbors. We release the dataset of 9 content-based and 3 collaborative-filtering latent representations of 29 275 tracks from Jamendo that we use for the evaluation. Moreover, we perform an online user experiment to compare the perceived track-to-track similarity of the selected evaluated latent spaces. The results show that content-based spaces show better results in our scenario, particularly embeddings from penultimate layers of auto-tagging architectures.