Problem
Booking.com has many surfaces that show recommendations: destinations, hotels, flights and attractions on the Homepage and Search Results pages. Each team needs its own logic. A model trained on data from a fixed period gives good but imperfect predictions. Availability, safety incidents, strikes and legal limits change at any moment, so a raw prediction can point to a place the customer cannot book. Without a shared structure each team rebuilds the same filtering, enrichment and sorting steps.
Example from the article: a destination carousel must start from the origin country for a new visitor but from recent searches for a returning one, and it must drop blacklisted destinations and destinations with no available properties in the customer's booking window.
Solution in brief
The Recommendation Platform (RecP) runs every request through one fixed pipeline: Process, Predict, Initial Filtering, Enrich, Enriched Filtering, Select & Sort. A team defines a recommendation strategy by picking components from a shared registry: models (combined into a prediction graph), features, filters, a selector and a sorter. A new use case becomes a configuration.
Questions this case answers
- How to let independent teams assemble recommendation strategies from shared models, features and rules without duplicating serving code
- How to apply real-time business constraints to model predictions, so that stale training data does not surface unavailable or forbidden items
- How to let many product teams launch their own recommendation carousels without rebuilding the serving logic each time?
- How to combine ML predictions with real-time business rules such as blacklisted locations and availability?
Tools and methods
- Recommendation pipeline — six fixed stages for every use case
- Prediction graph — models in sequence or in parallel, with a fallback model
- Feature mapping — model arguments bound to named features
- Components registry — shared catalog of components and use cases
- Rule-based filtering — real-time rules applied before and after enrichment
Google or a one-time email code. No password.