Chapter 7: ML system design and experiments
Candidate generation: two-tower and matrix factorization
Every video, shop or music app must find a short list of good items in a huge catalog before the page loads. Recommender design interviews ask exactly this. They ask how you get the short list, how you train its model and how you serve it fast.
What you will learn
- Explain matrix factorization from implicit feedback
- Train a two-tower model with in-batch softmax and explain sampling bias correction
- Serve candidates with an ANN index
Key terms
- candidate generation
- the first stage that picks a short list of items from a huge catalog
- embedding
- a list of numbers that places a thing in a space; similar things sit close
- two-tower model
- two networks that each turn a user or an item into a vector
- ANN index
- a structure that finds near-closest vectors fast, at a small cost in accuracy
Prerequisites
Take Ads click prediction: pCTR, calibration and expected value first. It comes right before this lesson in the course.
Where this lesson picks up: You built an ad ranker that scores each ad with care, one at a time, by calibrated pCTR × bid. A careful score for every item is too slow for a huge catalog, so how does a system choose which items to score?
What comes next
With a short list of good candidates in hand, in what order do you show them?
Next lesson: Ranking and re-ranking: multi-objective, diversity, cold start
More lessons in this part
Take the lesson
Short cards, one idea at a time, with quick checks as you go.