The recommendation
algorithm does not know you.
It just runs mathematics on
what you do.
There’s a common belief
that Instagram, YouTube, and
Spotify somehow “know”
you, that some intelligent system
is reading your mind, and
watching your every move.
It’s wrong.
What these platforms run
are recommendation algorithms. They
look at what you do on the app, compare
it with what millions of other users do,
and make a guess about what you might
want to see next. Sometimes the guess is
good; sometimes it isn’t. No magic. Just
mathematics.
Platforms have millions of videos,
songs, and posts. You can’t scroll
through all of that. So, the algorithm’s
job is simple: out of everything available,
pick the few things most likely to keep
you watching, listening, or scrolling.
There are two standard ways these
systems do this.
The first is to look at what similar users
liked – a method called collaborative
filtering. For example, say you and another
user have both liked the same ten
videos. That other user has also liked
an eleventh video you haven’t seen. The
system assumes you’ll probably like it
too, and recommends it to you.
The algorithm does not know if the
video is a cooking tutorial or a cricket
highlight. It only cares about patterns:
who liked what, who watched what,
and who skipped what. If your pattern
matches someone else’s, their likes become
your suggestions.
To understand how this works, imagine
a spreadsheet. Every row is a user,
and every column is a video or song.
Each cell shows whether that user liked,
watched, or skipped it. Most of the
spreadsheet is empty, because no one
person watches everything.
The algorithm guesses what
should go in the empty cells,
based on the patterns it sees
in the filled ones.
The second method is
to look at the content itself,
known as content-based
filtering. Instead of asking,
“What did similar users
like?”, it asks, “What does this
content look like, and does it
match what this user usually watches?”
On Spotify, this means measuring the
actual properties of a song. If you keep
playing fast, upbeat songs, it will recommend
more fast, upbeat songs. On
YouTube, the system looks at the video’s
title, description, tags, and category.
Newer systems can “watch” the video
themselves and “listen” to the audio, to
understand what the video is about.
No major platform uses just one method.
In practice, they combine all of them,
plus extra information like what time of
day it is, what device you’re using, and
how long you’ve been on the app.
The newer development is generative
AI being added on top of these systems.
Spotify now lets you describe a mood
in plain words, and it builds a playlist
around that description. YouTube uses
AI to auto-generate summaries of videos
so it can better understand what they’re
about. These additions make the recommendations
more accurate, but the basic
logic underneath—filter, then rank—is
still the same.
Long story short: a recommendation
algorithm is a filtering and ranking system.
It takes data about you and runs it
through models built on how millions
of other users behaved. It does not know
your name, or understand why you like
what you like. It merely spots patterns,
and, when the pattern match is good, the
suggestion feels personal. When it isn’t,
you scroll past—and that scroll becomes
another data point for the next round.