Stage 2 · Intermediate · M6
What to Show Next
About 80% of the hours people stream on Netflix come from what its recommendations put in front of them, not from a search.
14 lessons · 131 minSteady
About this chapter
Your screen shows five things and the catalog holds millions, so something has to choose. This chapter builds that chooser from a small table of people and dishes that is mostly blanks. You will fill the blanks with the people most like you, then with two thin tables of taste numbers, then with a map where people and things live together. Then the hard part: clicks that lie, scores that shine offline and flop in a real test, and a feed that slowly teaches you what it already believed.
What you will be able to do
- 18 min
Too Much to Show
Say what a recommender does, and why it is a machine for filling holes in a table.
- 29 min
Near Tastes
Measure how alike two people are from the things both rated, and why each person's average has to come off first.
- 310 min
People Like You
Fill a blank with a vote from the people most like you, and say why shops compare items instead.
- 49 min
Two Thin Tables
Fill every hole with the product of a person's numbers and a thing's numbers, fitted to the filled cells only.
- 511 min
Fitting Only What You Know
Write the loss a factor model minimizes, run the nudges that shrink it, and say why blanks must never count as zeros.
- 69 min
A Map of Taste
Read people and things as arrows on one map, score a pair with a dot product, and find what sits nearby.
- 79 min
The Dish Nobody Has Rated
Say why a new item or a new person breaks a ratings model, and fill the gap from what the thing is.
- 89 min
A Click Is Not a Like
Read taps, minutes and orders for what they are, and name the two ways they mislead: silence and position.
- 99 min
Find, Then Rank
Split a recommender into a fast first stage that finds candidates and a careful second stage that orders them.
- 1010 min
Scoring a List
Test a feed offline on a hidden week, and compute precision and recall at k.
- 119 min
Offline Wins, Online Losses
Say why a better offline score can lose a live test, and read a live result before shipping.
- 1210 min
The Feed That Feeds Itself
Show how a feed trained on its own output narrows, and how a little exploring breaks the loop.
- 1310 min
Varied, New and Fair
Judge a list as a whole, re-rank it for variety, and say what fairness means for the people who make things.
- 149 min
One Space for Search and Suggestions
Treat search and suggestions as one move, nearest arrows to a starting arrow, and say how it scales to millions.