Why implicit rather than explicit
Netflix and Amazon settled one question long ago: people watch films and buy products readily, and
rate them rarely. Studies put the share of users who leave explicit ratings below 1%. No
recommendation model can be trained on that.
Implicit signals solve the coverage problem: every visit, every click, every view is data. There is
several orders of magnitude more of it than explicit ratings, and it asks nothing of the shopper.
The signal hierarchy in e-commerce
Not every implicit signal carries the same weight. A practical hierarchy by strength:
| Signal | Interpretation | Weight |
|---|---|---|
| Purchase | Maximum interest | ★★★★★ |
| Add to cart | Strong interest | ★★★★☆ |
| Wishlist add | Interest without readiness to buy | ★★★☆☆ |
| Long dwell (15 s or more) | Studying the item | ★★★☆☆ |
| Click in a listing | Baseline interest | ★★☆☆☆ |
| Product page view | Weak signal | ★☆☆☆☆ |
Recommendation engines assign numeric weights to these signals and aggregate them into a profile.
The noise problem
The main weakness of implicit data is that it never says I do not like this. A non-click is treated
as a weak negative signal, which is imprecise: the shopper may simply not have seen the item because
it sat at the bottom of the page.
Good implementations account for position bias — items in the first positions get clicked more
often purely because they are visible. That does not make them five times more interesting than the
item in fifth place. Models with position correction (unbiased learning to rank) produce better
recommendations.
Tip: when collecting implicit data, always log the impression context alongside the event —
position, strategy, page. Without that context the signal cannot be weighted correctly at training
time.
Use in session-based recommendations
The implicit signals of the current session make recommendations possible with no history at all,
which is the answer for anonymous visitors and for cold start. A shopper views three pairs of
sneakers, the system infers the category interest and recommends similar items. It happens in real
time, within seconds, with no sign-in.
Common mistakes with implicit data
- Equal weight for every event: a purchase and a view are fundamentally different signals, and
treating them alike produces poor recommendations - Ignoring position bias: a high click-through rate in the top slots is not the same as relevance
- Duplicate events: one view logged several times on a page reload distorts the profile without
deduplication - A data window that is too short: a week of signals paints a different picture than 90 days, so
the horizon has to be chosen deliberately