Collaborative filtering

Collaborative filtering builds a matrix of users against items, filled with interactions such as purchases, clicks or ratings, and looks for structure in it. User-based variants find people whose histories overlap with yours and recommend what they chose; item-based variants find items that are frequently chosen together and recommend neighbours of what you already have. Matrix factorisation and modern neural approaches learn compact embeddings for users and items from the same signal.

Its appeal is that it needs no item descriptions: it never has to know that two products are both blue jackets, only that the same people buy them. That makes it strong for large catalogues with sparse metadata and for surfacing unexpected but relevant items.

Its weakness is the cold start problem. A new user has no history to match and a new item has no interactions to learn from, so pure collaborative filtering says nothing about either. Content-based methods, which compare item attributes, fill the gap, and most production recommendation engines blend the two. It is not the same as "customers also bought" counting, though that simple co-occurrence is a useful baseline.

• Product manager: it explains why recommendations feel eerily good for regulars and useless for newcomers, and why onboarding signals matter.
• Data lead: interaction volume and sparsity determine whether it works; a catalogue where most items have a handful of purchases needs hybrid methods.
• CTO / Head of Engineering: the models are cheap to train and serve, which makes it the right first algorithm to test against your rule-based baseline.
• Founder / CEO: it is the technique behind most of the personalisation you have experienced as a consumer, and it is entirely achievable on your own data.

Collaborative filtering exists because describing every item well enough to compare them is expensive and often impossible, while behaviour is recorded for free. People's choices carry taste that no attribute schema captures, and pooling those choices across users lets a system recommend without understanding. The trade-off is that it only works where behaviour exists: it is silent on new users and new items, it can amplify popularity, and it can be gamed by fake interactions. Used as one component of a hybrid engine, with content-based and rule-based fallbacks, it is dependable.

• Item-to-item recommendations on a fashion retailer's product page, learned from co-purchase across all customers.
• Suggesting courses to a learner from what students on similar tracks completed and rated highly.
• Recommending SaaS integrations to an account from adoption patterns of comparable workspaces.
• Surfacing restaurants or dishes in a delivery app based on customers with overlapping order histories.
• Proposing mutual funds or savings products to bank customers with similar transaction profiles, within suitability rules.

Technique / practiceA modelling technique, one of several inside a recommendation engine. Eazyware selects and benchmarks it against content-based and hybrid approaches during a personalisation engines build, with results judged by controlled test.

Eazyware service that covers it: Personalization Engines. Starting prices are on the pricing page.