...make recommendations
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Recommendation ... Embedding ... ...cluster ... ...optimize decisions ... ...answer questions from my data
Use this page when you want to show each person the items they are most likely to want: products, videos, articles, jobs, songs. Recommenders mix prediction ("will they click?") with search ("which few thousand items are worth scoring?").
Start here
- Do you have interaction history (views, clicks, purchases, ratings)?
- No (cold start) → popularity plus business rules; content-based similarity using item embeddings made from descriptions or images; a short onboarding question
- Yes → continue
- How big is the catalog and the traffic?
- Small or medium → item-to-item collaborative filtering ("people who bought X also bought Y") or matrix factorization (e.g., implicit ALS)
- Large → two-stage system: retrieve a few hundred candidates (two-tower embeddings + vector search), rank them with a gradient-boosted or deep ranking model, then re-rank for diversity, freshness and business rules
- Does order or recency matter (next video, next song, this session)?
- Yes → sequential recommenders (SASRec, BERT4Rec) or generative recommenders: transformers trained on interaction sequences, the way LLMs are trained on words
- Do you need to keep exploring new items and adapt in real time? → contextual bandits: ...optimize decisions
- Do users want to ask in their own words, or need explanations? → an LLM with retrieval over your catalog: ...answer questions from my data
- Buy instead of build? → managed services such as Amazon Personalize, or commerce search platforms; open source: Microsoft Recommenders, TorchRec
Quick picks
| Situation | Try first | Then compare |
|---|---|---|
| New site, little data | Popular items + content similarity | Simple rules by category |
| Purchases or views available | Item-to-item collaborative filtering | Matrix factorization |
| Millions of users and items | Two-tower retrieval + gradient-boosted ranker | Deep ranking models |
| Sessions and sequences | SASRec / BERT4Rec | Generative recommender |
| Many new items, fast feedback | Contextual bandit | Exploration slots in the ranked list |
Real-world examples
- Amazon — item-to-item collaborative filtering, which scales because it compares items rather than users, has powered Amazon's "customers who bought this also bought" recommendations for over two decades. The history of Amazon's recommendation algorithm | Amazon Science
- Netflix — Netflix reported that recommendations drive about 80% of the hours members stream. The Netflix Recommender System | C. Gomez-Uribe & N. Hunt - ACM TMIS In 2025 it described replacing many specialized models with one transformer "foundation model" trained on members' full interaction histories, served to other systems as embeddings or as part of their models. Netflix foundation model for personalized recommendation | ZenML LLMOps Database
Watch out
- Feedback loops — users can only click what you show; reserve some exploration.
- Popularity bias and filter bubbles — measure diversity and coverage, not only clicks.
- Offline metrics often disagree with online results — confirm with A/B tests.
Measure success
Offline: precision@k, recall@k, NDCG, coverage, diversity, novelty. Online: click-through, conversion, watch time, retention, measured with A/B tests. More: ...evaluate and trust AI
Learn more
- Recommendation systems course | Google for Developers
- Recommenders: best practices and examples | Microsoft / Linux Foundation AI
- Recommendation ... Embedding ... Database; Vector & Relational
Where next? ◀ Predict & analyze · ...cluster · ...optimize decisions · ...answer questions from my data · ▶ ...evaluate and trust AI