Skip to main content

Recommender systems

Recommending is neither predicting a star rating nor searching a corpus: it is ordering a catalog of thousands of items for every single person, from very few signals and a great deal of bias.

Course Duration: 6h​

What You'll Learn​

  • Frame a recommendation task as a ranking problem, not as a rating regression
  • Build user-based and item-based collaborative filtering from a sparse rating matrix
  • Learn latent factors with matrix factorization (SVD, ALS) and regularize them properly
  • Represent items by their content with embeddings of course descriptions
  • Combine content and collaborative filtering into robust hybrid systems
  • Train a two-tower neural retriever in PyTorch with negative sampling
  • Handle the cold start for new users and new items with content and controlled exploration
  • Evaluate an engine offline with recall@k, NDCG@k, coverage and diversity
  • Correct the position bias and popularity bias hidden in implicit feedback
  • Deliver an evaluated recommendation engine with an honest offline-to-online story

Prerequisites​

  • Course 03 — Mathematics for AI (linear algebra, gradients)
  • Course 05 — Unsupervised learning (similarity, dimensionality reduction)
  • Python, NumPy and pandas; PyTorch for module 6

The thread running through the course​

Every module anchors in the catalog of an online learning platform modeled on InSkillML: about 500 courses, tens of thousands of learners, a sparse matrix of explicit ratings (stars) and of implicit feedback (enrollments, minutes watched, completions). Each module adds one piece to the engine, from the simplest neighborhood method to a two-tower retriever. The public MovieLens dataset serves as a checkpoint whenever a reference number is useful.

Course Modules​

  1. Framing a recommendation problem
  2. User-based and item-based collaborative filtering
  3. Matrix factorization and SVD
  4. Content-based filtering and similarity
  5. Hybrid approaches
  6. Deep recommendation and embeddings
  7. Cold start: new users and new items
  8. Metrics: recall, NDCG, coverage, diversity
  9. Implicit feedback and position bias
  10. Project: an evaluated recommendation engine

Assessment and certificate​

The course ends with a 40-question exam covering the ten modules, from diagnosing a mis-framed problem to reading a recall-coverage curve. On success, a certificate of completion is issued immediately; its number is verifiable on the platform by any third party.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.