Problem
LinkedIn Premium has four products with different features and prices. The team has to decide which product to promote to each member, with which message, and in which channel (email, feed, notifications, job search, homepage). A wrong time, place or message annoys members and wastes the chance to help them. Testing creatives by hand took a lot of manual effort and delivered little personalization.
Example from the article: a member sees a Premium message on the homepage or in a "who viewed your profile" surface. The system must choose the product, the creative and the channel, or show nothing.
Solution in brief
Treat every (product, creative, channel) triplet as an arm of a contextual bandit. A two-tower neural network scores member and arm, and a Thompson sampling layer on top adds exploration. New arms are scored from their attributes, so no retraining is needed.
Questions this case answers
- How to pick the best offer, message and channel for each user when the set of combinations is in the thousands and sequential A/B tests are too slow
- How to score newly added candidates in a contextual bandit without retraining the model
- How to choose which product, creative and channel to show each user, and when to show nothing?
- How to run hundreds of creative tests in weeks instead of slow sequential A/B tests?
Tools and methods
- Contextual multi-armed bandit — chooses one arm per member and keeps collecting reward information
- Two-tower model — separate member and arm embeddings, precomputed offline for fast scoring
- Neural Thompson Sampling (last-layer approximation) — makes the score stochastic so the model explores
Google or a one-time email code. No password.