A Recommendation System ranks items for a user by predicted relevance. The three classical families: collaborative filtering (learn from how similar users behaved), content-based (learn from item features and user history), and hybrid (combine both, usually in a two-stage retrieval-then-ranking pipeline). Modern production systems lean heavily on learned embeddings for candidate generation and gradient-boosted or neural rankers for final ordering.
Cold-start is the persistent failure mode: new users and new items have no interaction history, so the system falls back to content features or popularity until enough signal accrues. Most teams mitigate with explicit onboarding prompts, editorial defaults for the first session, and contextual bandits that explore just enough to learn.
Eval is trickier than classification accuracy. Offline metrics (NDCG, MAP, HitRate) correlate imperfectly with online KPIs (engagement, conversion, revenue per session). Production teams run A/B tests on every meaningful change and watch for diversity collapse — the model that maximizes click-through often narrows the catalog exposure and hurts long-term retention.
Where AIMenta applies this
Service lines where this concept becomes a deliverable for clients.
Beyond this term
Where this concept ships in practice.
Encyclopedia entries name the moving parts. The links below show where AIMenta turns these concepts into engagements — across service pillars, industry verticals, and Asian markets.
Other service pillars
By industry