What is shown shapes behaviour, which then reshapes what is shown

What it is

A recommender does not first possess a fixed taste and then serve it. What appears on screen decides what the user clicks; what is clicked decides what appears next. The display-behaviour loop is that closed coupling: the display set shapes observable behaviour, and behaviour is written into the next display. Neither side is exogenous.

Until that loop is opened, click logs cannot be read as “this is what the user already liked.” The log is a product of the loop.

Why it happens

Display is the sampling frame for behaviour. Items never ranked onto the screen get no chance to produce a positive act; items in the first slot, even if only slightly better, take a disproportionate share of clicks. Those clicks return as labels, and the next sort pushes similar items further up. Once the coupling is closed, input (display) and output (behaviour) define each other; the system is fitting its own last decision.

People in the loop are not neutral sensors. Position, autoplay, and “next episode” push behaviour toward what was already shown. “User choice” and “system choice” are tangled in the log and cannot be unpicked by treating clicks as preference.

Studying it

To see the loop you have to manipulate display. Hold ranking fixed for some sessions on one slice of traffic; on another, inject displays independent of the current model (random or editorial slots); then watch whether subsequent behaviour follows the display. Independent variables: whether the display policy is closed-loop, injection share. Dependent variables: category mix of later clicks, mutual information with the display policy, divergence between self-reported preference and logged preference.

Offline replay on historical logs cannot see the loop — the log is already the product of one display policy. That offline metrics favour the logging policy is a later problem. Here the question is whether the coupling itself exists.

Where it stops holding

One-shot queries (search a known title, open a shared link) barely loop; display is set by the query or the inbound link. On heavily edited timelines (live, a news front page) display is mostly human scheduling and the loop is weak. In catalogues with few users and many items the loop turns slowly and will not show in a few days. This entry only argues that display and behaviour write each other. It does not treat how early bias is amplified, and it does not treat the need to pour in a signal from outside the loop.

Applying it

• Store the current display policy as part of the behavioural record. Clicks without display context should not go straight in as preference labels.
• Keep a regular contrast rail that does not follow the current model (editorial slots, or a small random share) so you can see whether behaviour still diverges from the main rail.
• Check: in shadow traffic, drop a class from the first screen for a week. If clicks on that class vanish with it, clicks have been a function of display, not an independent taste.

Related

• Same group: L6.04.2 The loop amplifies whatever bias arrived first · L6.04.3 Breaking self-reinforcement needs a signal from outside the loop
• Nearby: L6.09 Feedback Loops and Preference Entrenchment · L6.02 Filter Bubbles · L3.13 User Feedback Loops on Generation Quality
• Search terms: display-behaviour loop · closed-loop ranking · feedback loop