Cold-Start Playlist Generation

Cold-start playlist generation is the task of automatically populating a playlist given only its title. This task presents a fundamental challenge in music information retrieval in the absence of user listening histories and due to the inherent vocabulary gap between playlist titles and track metadata. In this work we investigate multiple novel approaches, including transfer-based playlist initialization, multi-stage re-ranking, and a hybrid single-stage retriever that combines dense and lexical models. Our findings show the effectiveness of these new approaches and explores possible efficiency trade-offs. Additionally, we implement a new evaluation strategy to measure the musical feature overlap of playlists and retrieved tracks that captures semantic relevance instead of exact match of tracks.