Best modeling approach for audiobook recommendations, cross-content relationships, and metadata use

Develop and evaluate recommendation modeling approaches for audiobooks that (i) effectively represent audiobook content, (ii) capture relationships between audiobooks and other audio content types such as podcasts and music, and (iii) systematically exploit available metadata to improve recommendation quality at scale.

Background

The introduction highlights the novelty and challenges of deploying audiobook recommendations at scale on a platform already optimized for music and podcasts. Because audiobooks were initially paywalled and interactions are sparse, the choice of modeling approaches and metadata utilization has heightened importance.

The authors propose 2T-HGNN as one solution, but explicitly note that the broader question of how to best model audiobook content, its relationships with other content, and metadata usage had not been settled at the time of writing.