Recommender systems
What should this person see next? Collaborative filtering and matrix factorisation, embeddings of users and items, the two-stage retrieve-then-rank architecture, ranking metrics — and the feedback loop that makes recommendation unlike any other prediction problem.
50 min
After this lesson you can
- Distinguish explicit from implicit feedback and say what each can and cannot tell you
- Write the biased matrix-factorisation model and its regularised objective, and fit it by alternating least squares
- Explain users and items as embeddings, and why dot-product and cosine retrieval behave differently
- Describe the two-stage retrieval-and-ranking architecture and the two-tower model
- Compute recall@k and NDCG@k by hand
- Explain feedback loops, popularity bias and cold start, and why offline metrics do not settle them
In this lesson
- The data: an enormous, nearly empty matrix
- Collaborative filtering: people like you liked this
- Matrix factorisation: a few hidden tastes
- Embeddings and retrieval
- The architecture: retrieve, then rank
- How recommenders are evaluated
- Feedback loops, popularity bias and cold start
- Open problems
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.