Reproduce “Cold-Starting Podcast Ads and Promotions with Multi-Task Learning on Spotify” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shivam Verma, Hannes Karlbom, Yu Zhao, Nick Topping, Vivian Chen, Kieran Stanley, Bharath Rengarajan
Spotify developed a multi-task learning framework to unify podcast ad and promotion targeting, improving cold-start performance for less-streamed content and enhancing efficiency across both paid and organic content discovery. The model delivered up to a 60% increase in impression-to-stream rate and a 38% reduction in effective cost-per-stream for low-stream podcasts in A/B tests.
Key insights
- Unifying advertising and organic content promotion tasks within a single MTL model effectively leverages shared representations to mitigate the cold-start problem for new content and objectives.
- Strategic optimization techniques, such as adaptive loss masking and source-balanced sampling, are critical for managing knowledge transfer and addressing data imbalance in multi-task learning systems.
- Transfer learning, enabled by the MTL framework, allows data-rich promotion tasks to significantly benefit data-scarce ad tasks, yielding substantial cross-channel performance improvements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.