A social media app can't see what you think. It sees what you do: what you open, how long you stay on a post, what you like, comment on, share, follow, search for, and hide, and when you do it. Together these actions make up your data footprint. (You'll audit your own in 3.1.2.)
A platform then turns that footprint into a profile: a set of labels and scores that describe what you seem to like. Giving labels to a person from data is called profiling. It is the same idea as the classifier in 1.1.2, which put a label on an image, except the thing being labeled is you.
Every post carries tags, like #space or #baking. Here, each action gives points to every tag on the post, and your profile is just the running total for each tag. The numbers below are made up so you can do the math by hand.
| What you do | Points for every tag on that post |
|---|---|
| Tap like | +2 |
| Write a comment | +3 |
| Send it to someone | +4 |
| Tap "not interested" / hide | −4 |
Real platforms use many more signals, such as how long you watch, whether you replay something, and when you scroll past. They also learn the weights from data instead of using fixed numbers like these.
Rio is an invented user. In the first round of the feed lab, Rio shares the telescope post (#space), likes the farming speedrun (#gaming), hides the sourdough post (#baking), comments on the dog-trick post (#pets), and likes the sunrise ridge post (#hiking).
Your actions are facts about what you did. A profile goes further: it says what you are "interested in" or "like." Those are conclusions, and conclusions can be wrong. Try sorting claims about an invented user.
Label each statement: In the data (the platform recorded it), Reasonable guess (strong evidence, still a guess), or Stretch (the data can't support it).
Correct so far: 0 / 10 · answers tried: 0
A recommender decides what to show you next. A simple one gives every post a score and puts the highest scores first. Here the score is the sum of your profile points for the tags on the post. A post tagged #space and #gaming adds both. A tag at −4 pulls a post down.
Rio's profile from the last example: space 4, pets 3, gaming 2, hiking 2, baking −4. Three unseen posts could go in the next feed.
Build a profile and watch the feed respond. Use your own taste, without any names. After each round, check one post's score by hand. Then change the settings and see what happens to the variety of topics.
Round 1 shows one post from each of 8 topics. React to the ones you'd really react to. Leave the rest alone.
Top interests it would guess: none yet
These are guesses about what you would click on, not facts about you.
React to at least one post to build the next feed.
Compare with a partner: don't use names. Compare your "top interests" lists and the topics in your feeds. Which of your reactions made the biggest difference?
Real feeds work the same way in outline, with much more going on: millions of posts, hundreds of signals, and models that learn the weights. New posts get labels too, sometimes from the people who post them and sometimes from systems that learned to label content from examples. That is why a single like can change what you see so quickly.
Your reactions build the profile. The profile ranks the feed. The feed is what you react to next. That is a loop, and it can feed on itself: the more you engage with one kind of post, the more of it you see, and the less you see of everything else.
Here is an invented user who reacts the same way every time: they like any post about space or gaming and ignore everything else. After round 1, the profile has points on only those two tags. Here is what the next feed of 6 posts looks like with different numbers of explore slots, which are slots kept for topics the profile hasn't scored.
| Explore slots | Posts matching their interests | Topics in the feed |
|---|---|---|
| 0 | 6 of 6 | 3 of 8 |
| 2 | 4 of 6 | 6 of 8 |
| 3 | 3 of 6 | 6 of 8 |
With no explore slots, all 6 posts match and the feed covers 3 of 8 topics. Explore slots widen the feed, and they cost a little matching. This pool has only 40 posts, so it runs out of matches quickly. A real feed doesn't run out, so the narrowing can continue.
The score is a design choice. A company can build a recommender to maximize time on the app, clicks, shares, purchases, or something else, such as variety or a user's own stated goals. Feeds are very often tuned to engagement, because engagement is easy to measure and can be tied to money. Reasonable people disagree about how much of a feed should be tuned that way, and about who should decide.