Recommendation system

A recommendation system predicts which items, products, articles, videos, connections, a specific user is most likely to want, and ranks them to decide what to surface, on a homepage, in an email, or as a "customers also bought" module. It differs from search, which responds to an explicit query; recommendations are typically generated proactively from a user's history and context.

Two families of technique dominate: collaborative approaches, which recommend items liked by similar users without needing to understand item content, and content-based approaches, which match items to a user based on their attributes, similar to product affinity analysis. Most production systems blend both, along with business rules such as inventory limits or promotional priorities, and increasingly use a propensity model to weight recommendations by predicted likelihood of action.

Recommendation systems matter because they drive a large share of engagement and revenue on content and e-commerce platforms, and they are closely related to next-best-action systems, which extend the same idea beyond product suggestions to any recommended action. Common pitfalls include feedback loops, where the system mainly recommends what it has already shown a narrowing set of customer segmentation groups, and the "cold start" problem of having too little data to recommend well for new users or new items.