Chapter 2 · Part 1
You are what you tap
To score items for you, the system needs to know your taste. But you never told it anything — no survey, no list of favorite genres. So where does the profile come from? You build it, constantly, without noticing. Every tap, pause, replay and skip is a tiny vote, and the feed is listening to all of them.
These are called implicit signals: feedback you give through behavior rather than by explicitly rating something. They're noisier than a thumbs-up, but there are thousands of them, and they're brutally honest.
Scroll to watch a blank profile fill in from a stream of ordinary actions.
A brand-new account knows nothing about you — the profile starts blank.
Explicit vs implicit feedback
- Explicit feedback is when you directly rate something: a star rating, a thumbs-up, a "not interested." Clear, but rare — most people almost never do it.
- Implicit feedback is inferred from behavior: clicks, watch-time, scroll depth, replays, saves, shares, and the negatives — fast skips, "hide this," quick bounces. Abundant, and available for every interaction.
Modern feeds run almost entirely on implicit signals, because there are so many of them. Watch-time in particular is gold: how long you stayed says far more than whether you clicked.
Signals are noisy — and that's fine
Any single action can mislead: you might watch a video to the end because you fell asleep, or click a thumbnail you immediately regret. Implicit feedback is full of this noise. It works anyway because of volume — across hundreds of interactions, the accidents average out and the real pattern shows through.
WEIGHTS = {"finish": +3, "like": +2, "comment": +2, "skip": -2}
profile = defaultdict(float)
for event in user_events:
profile[event.category] += WEIGHTS.get(event.action, 0)
# profile -> {"cooking": 11, "travel": 4, "sports": -2, ...}Where we're headed
Now we have a profile of what you like. But the most powerful trick in recommendations barely looks at your profile directly — instead it asks: who else behaves like you, and what did they love that you haven't seen yet? Next: collaborative filtering.