Reproduce “Multimedia Asset Personalization via Multimodal Embeddings at Netflix” 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
Emma Yanyang Kong, Aditya Deshpande, Bowei Yan, Asad Abbasi, Santiago Castro, Avneesh Saluja, David Fagnan, Ashish Rastogi
Netflix integrated multimodal embeddings from foundation models, including CLIP and an in-house tri-modal MediaFM, into its content personalization systems for artwork and video previews. This approach addressed cold-start challenges and demonstrated a +0.127% lift in the core discovery metric for artwork and a +0.193% lift in streaming hours for video previews.
Key insights
- Integrating content-rich multimodal embeddings into personalization models effectively mitigates the cold-start problem and facilitates knowledge transfer across diverse assets.
- Consolidating multiple personalization models into a unified architecture, particularly when combined with content embeddings, improves efficiency and boosts performance on data-starved content surfaces.
- Developing a lightweight offline proxy task, based on predicting unpersonalized popularity winners, provides a reliable and cost-effective method to screen candidate embedding models before extensive online A/B testing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.