Leveraging Artist Catalogs for Cold-Start Music Recommendation

A research team from institutions including the University of Amsterdam and Spotify has published a novel solution to a core problem in music streaming: the 'item cold-start.' This refers to the challenge of recommending newly released tracks that lack user interaction history, which traditional collaborative filtering (CF) models rely on. The paper, 'Leveraging Artist Catalogs for Cold-Start Music Recommendation,' introduces ACARec (Artist Catalog-Aware Recommender), an attention-based neural architecture. Instead of treating new tracks as completely unknown, ACARec frames them as 'semi-cold' by leveraging the rich collaborative signal from the artist's existing body of work. It generates CF embeddings for new songs by having an attention mechanism analyze patterns in the artist's previous catalog.

ACARec demonstrates a significant performance leap, more than doubling standard recommendation metrics like Recall and Normalized Discounted Cumulative Gain (NDCG) compared to models that only use audio or textual content features. The model shows particular strength in two critical areas: facilitating new artist discovery by accurately projecting their style onto the recommendation landscape, and providing a more precise estimation of a new track's potential popularity. This artist-aware approach fundamentally shifts the paradigm from treating artist data as just another input modality to recognizing the inherent hierarchy between artists and their items, a nuance previous methods often missed. The work has been accepted for presentation at the UMAP 2026 conference.