Auto-tagging system based on song’s latent representations for inferring contextual user information

Abstract

Currently in the field of Recommender Systems for the music domain, there is active research about approaches for inferring the user context. Moreover, in the Music Information Retrieval there have been great advances in the generation of latent representations of songs including approaches such as contrastive learning as pretrain strategy or other approaches related to Natural Language Modeling like codified audio language modeling (CALM). Such advances are especially useful for Music Information Retrieval discriminative tasks such as genre classification, key detection, emotion recognition and music tagging. This last task attracts the interest of music streaming services that seek to tag their catalogs, especially with tags related to the user's context as this has a great impact on their tastes and influences the developed recommender systems. These tags are usually provided by users on social networks and are frequently found only for popular songs in the catalog. However, recently added songs to the catalog or songs belonging to the long tail do not have these tags and the need to create new systems called auto-taggers capable of tagging these songs arises. This paper proposes an auto-tagging system and presents an evaluation of different multi-label classification approaches included in it for contextual label auto-tagging. These approaches use different latent representations of songs, employing a recent published dataset with user context tags. The results obtained from the case study conducted to evaluate the proposed system show a clear improvement in the classification metrics by using new latent representations compared to the use of simpler features in traditional state-of-the-art approaches.