Recommendation System Architecture Basics

Most recommendation systems are built from two complementary techniques: collaborative filtering, which looks at what similar users liked, and content-based filtering, which looks at the attributes of the items themselves. Understanding how these combine, and how to handle the cold-start problem, is the foundation for building a recommendation system that delivers real value without unnecessary complexity.

Most recommendation systems combine collaborative filtering, which looks at what similar users liked, with content-based filtering, which looks at the attributes of the items themselves.

A cold-start problem arises for new users or new products with no history, which is usually addressed with simple popularity-based recommendations until enough data accumulates.

We start clients with a simpler hybrid approach before investing in more advanced deep learning recommendation models, since the simpler system often captures most of the value at a fraction of the cost.

Collaborative filtering works well once there is enough interaction history, since it can surface non-obvious connections, like two products frequently bought together for reasons that are not apparent from their attributes alone. Content-based filtering, by contrast, works from day one for any item with known attributes, and it explains its own recommendations naturally, since "similar because it shares these attributes" is easy to communicate to a user.

The cold-start problem cuts both ways: new users with no history and new items with no interactions both need a fallback strategy. Popularity-based defaults, trending items, or attribute-based similarity to items a new user has already viewed all provide reasonable recommendations until enough behavioural data accumulates to power collaborative filtering properly.

We generally recommend starting with a hybrid system combining both techniques with simple, explainable logic, and measuring its actual impact on engagement or conversion, before investing in more advanced deep learning recommendation models. In our experience, the simpler system captures most of the achievable value, and the case for the added complexity of a deep learning approach only becomes clear once that baseline's limits are actually reached.