Two kinds of signal
Explicit feedback is when a user deliberately rates something: five stars, a thumbs up, a like. It is clean but rare, since most people never rate.
Implicit feedback is behavior collected silently: clicks, watch time, purchases, dwell, skips. It is abundant but noisy, since a click can be a misclick and a long watch can mean the laptop was left running.
The hard part of implicit data
With explicit ratings you know what a user disliked. With implicit data you only see positives. A missing interaction is ambiguous: maybe the user disliked the item, or maybe they never saw it. You cannot treat all unobserved entries as negatives.
How models adapt
- Frame the value as a preference of one or zero plus a confidence weight that grows with how strong the signal is, as in weighted matrix factorization.
- Use negative sampling: draw random unseen items as likely negatives rather than labeling every blank.
- Optimize a ranking objective so seen items outrank sampled unseen ones.
Key idea
Explicit feedback is clean but scarce; implicit feedback is plentiful but only positive, so models use confidence weights and negative sampling.