5.22 Recommender systems
You can build a recommender that beats 'most popular'.
Before:03. Data Handling & AnalysisUnlocks:06. Deep Learning10. Production & MLOps13. Capstones, Portfolio & Interviews
Recommenders drive feeds, stores and streaming, built from collaborative filtering, content features or both — with implicit feedback, clicks rather than ratings, as the realistic regime with its own losses and ranking metrics. It closes the classical module as the applied composite. The two hard problems are named honestly: cold start, because new users and items have no history, and popularity bias, because recommending the popular is accurate, safe, and useless.
Work through these
Collaborative filtering: user-, item-, matrix factorization
Recommending from patterns of who liked what, either by finding similar users, similar items, or by factorising the whole interaction matrix. The third is the classical strong approach.
Content-based and hybrid approaches
Recommending from the properties of items themselves, and combining both approaches. Hybrids are what production systems actually use.
Implicit feedback and ranking metrics
Most real signals are clicks and views rather than ratings, which changes both the model and how it should be scored. Ranking measures matter more than error measures here.
Cold start and popularity bias
New users and new items have no history, and popular items crowd out everything else in the feedback loop. Both problems are structural and both need deliberate handling.
Sign in to keep your progress.
Free resources
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.