Module 1The recommendation problem
3 lessonsThe recommendation problem. What a recommender decides on the product page, the home page and the email, what success means in money, and the setup of the project with the synthetic shop log.
Show each customer the products that matter to them. Split the shop log by time, beat a strong popularity baseline, compare neighbours, matrix factorisation, content and a two-stage ranker on the same customers, and run the winner with business rules, a fallback and privacy controls.
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
You are enrolled. From now on, the lessons that you complete are saved.
This course teaches recommendation as a business system: implicit feedback and an honest time split, popularity and co-occurrence baselines, item-item collaborative filtering, matrix factorisation with ALS and BPR, content-based and hybrid methods, cold start, two-stage candidate generation and ranking, offline metrics by segment, simulated online tests with guardrails, and a service with business rules, a fallback and privacy controls. It does not teach supervised learning or gradient boosting from the start, vector search for documents, or how to design and analyse an A/B test in depth.
Take another course if you want to:
By the end of this course, you will be able to:
Larkfield, an online shop for home and garden products, shows the same twenty bestsellers to everyone. A customer who bought a dog bed last week sees garden chairs. A customer who opens the shop in July sees winter coats. Products added last month never appear at all. Larkfield wants better suggestions on its product pages, a personal home page and a weekly email, for returning customers and for new ones with no history.
In this course, you follow Priya, Larkfield’s data scientist, Omar, the product manager, Amira, the developer who owns the service, and Rafael, who sends the weekly email. You start with the data: views, baskets and orders as implicit feedback, an honest split by time, and a strong popularity baseline. Then you add one family of methods at a time and measure it on the same customers, in one comparison table.
You go from “people also bought” and item-item collaborative filtering to matrix factorisation with ALS and BPR, content-based suggestions for new products, hybrids and a two-stage system with a LightGBM ranker and business rules. You learn why the best offline score can lose an online test, and how a recommender can trap itself in a feedback loop.
You build on one project through the course: larkfield-recs, with a recs command, an evaluation harness and a small recommendation service with a fallback. The shop log is synthetic, made by our code with a documented model of customer behaviour, so you can see effects that real data hides. The case studies use real open data.
You completed Machine Learning, From Problem to Reliable Model, or you have the same skills (train, validation and test sets, leakage, gradient boosting, precision and recall). You can read and change Python and pandas code, use a notebook and a terminal, and read SQL results. The SQL and RAG courses help but are not required; lessons that use their ideas explain them briefly and link to them. You need Python 3.12 on your computer, or a free Google or Kaggle account for the notebooks. You need no key, no graphics card and no money.
Courses to take first: Machine Learning: From Problem to Reliable Model.
ML and AI engineers, data scientists and backend developers who build "you may also like", personal home pages or personalised emails, and product people who must decide whether personalisation is worth its cost.
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
The recommendation problem. What a recommender decides on the product page, the home page and the email, what success means in money, and the setup of the project with the synthetic shop log.
Data, implicit feedback and an honest split. Views, baskets and orders as weighted interactions, the sparse user-item matrix, a time split and what a random split hides, and three popularity baselines.
Measure recommendations offline. Precision, recall, MAP and NDCG by hand, coverage, novelty and diversity, and results for new, occasional and loyal customers.
Neighbours and collaborative filtering. "People also bought" with lift and cosine, item-item collaborative filtering with an explanation for each suggestion, and why user-user struggles in a shop.
Matrix factorisation for implicit feedback. Hidden tastes by hand, ALS and BPR with the implicit library, and tuning on a validation period without fooling yourself.
Content, hybrids and cold start. Product embeddings, new products and new customers, and hybrids measured by what they cost known customers.
Two stages, candidates and ranking. Several candidate sources, a LightGBM LambdaMART ranker, business rules and diversity with the cost of each, and a short look at sessions and LLMs.
From offline scores to the real shop. Position bias and biased logs, simulated A/B tests with guardrails, and feedback loops with a small share of exploration.
Run it in production. Batch or real-time serving with a popularity fallback, monitoring and retraining by segment, and deletion requests and sensitive products.
End-to-end case studies on real open data. Real retail transactions, browsing sessions and book ratings, and an optional "what to learn next" case. Complete at least one for the certificate; each one is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Recommend for an online retailer2 hr 30 min
On Online Retail II (real transactions, CC BY 4.0): popularity, co-occurrence, ALS and a two-stage ranker with a time split and customer segments.
Next item from browsing sessions2 hr 30 min
On a sample of the OTTO sessions dataset (clicks, carts and orders, CC BY 4.0): session-based suggestions and cold start for anonymous visitors.
Books for readers2 hr 30 min
On Goodbooks-10k (downloaded from its source, CC BY-SA 4.0): content-based, collaborative and hybrid methods with coverage and novelty against a popularity baseline.
Optional. Suggest the next lesson on a learning platform from synthetic progress data, where order and prerequisites matter more than popularity.
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.