Problem
A streaming service picks one image or video preview per title for each member. Models that treat every asset as an opaque ID have no signal for a new title, so the service falls back to popularity until interactions accumulate. Artwork also comes in five canvas crops, each with its own model and its own cold start. A TV home-screen redesign was about to send the most impressions to the canvas with the least history.
Example from the article: a member who engaged with a comedian's past stand-up artwork gets the new title's asset featuring that comedian, though the image was never shown before.
Solution in brief
Add a pretrained content embedding of each asset (CLIP for images, a multimodal model for video) to the ID embedding, so a new asset is understood at once. Train one model over all canvases, weighting examples by long-term reward. Screen embeddings with a linear probe and judge models offline with inverse propensity scoring (IPS) on exploration traffic.
Questions this case answers
- How to personalize a newly added item whose interaction history is empty, when a standard ID-based model treats it as unseen?
- How to pool training signal across several data-poor serving surfaces without hand-tuned per-surface weights?
- How to screen many candidate embeddings cheaply before running end-to-end online experiments?
- How to personalize artwork or video previews for a new title that has no interaction history yet?
Tools and methods
- CLIP — 768-dimensional image embedding of each artwork; its text side scores query-to-artwork match.
- Reward-based example weighting — weights training examples by the long-term reward of their interaction type.
- Inverse propensity scoring on exploration traffic — unbiased offline estimate of a candidate policy's reward.
- MediaFM, SeqCLIP — video preview embeddings.
- Linear probe — predicts the debiased popularity winner from the embedding alone.
Google or a one-time email code. No password.