Feedback Loops, Popularity Bias, and Ecosystem Dynamics

A policy allocates exposure. Exposure changes interactions. Interactions train the next policy. Popularity, familiarity, provider investment and catalog survival can therefore be outcomes of recommendation as well as inputs.

Feedback loops are not always harmful. They can help good items accumulate evidence. The governance question is whether early noise, privilege or narrow objectives become self-reinforcing faster than the system can correct them. That is a question about rate, and rate has to be measured rather than argued.

It has been measured, on MovieLens. Between February 2008 and August 2010, 1,405 users supplied 173,010 ratings on 10,560 distinct movies and opened their "Top Picks For You" page 150,759 times. The average pairwise content diversity of the top-15 recommendations fell from 25.02 to 24.67 (t-test p = 2.43e-06). The diversity of the movies users actually rated fell from 26.60 to 26.01 (p = 1.542e-12). "We do find that recommender systems expose users to a slightly narrowing set of items over time," the team that ran the study wrote, in 2014.

Read the second half of their result before deciding what the first half means. The drop was smaller for the users who followed the recommendations (26.67 to 26.30) than for those who ignored them (26.59 to 25.86). The narrowing was real, statistically unmistakable, small, and worse among the people who refused the recommendations than among the people who took them. That is what the rate question looks like when somebody answers it. A signed, sized number on a named population — not a verdict for or against loops.

A follow-up experiment took the same 48 songs back to market and lied to the participants about them. Between 14 March and 10 August 2005, 12,207 people passed through: 2,211 before the intervention, 9,996 after. At the switch, the displayed download counts were artificially inverted. The song presented as the most downloaded was the one that had in fact been downloaded least. The songs did not change. Only the record of what other people had supposedly done changed.

The falsified popularity largely made itself true. "Overall, the final rankings of almost all songs seem to be permanently affected by the inversion, where songs that were promoted by the inversion tended to do better in the long run, and songs that were initially demoted tended to do worse," Salganik and Watts reported in 2008. One lie, told once, at one moment. The ranking carried it from there without further help.

The cycle has no natural end. An initial state — catalog quality, prior exposure, editorial choices and cold-start priors — is converted into policy exposure, as retrieval and ranking allocate the visible positions. A behavioral response follows: users click, consume, ignore, report or leave. Then a provider response: creators and sellers adapt content, price, inventory and strategy. Retraining and path dependence close the ring. New models learn from the altered interaction and supply distribution, so the catalog quality and cold-start priors of the initial state keep reappearing years later.

That last stage has been turned into a measurement instead of a label. A 2020 simulation ran user interaction through several recommendation algorithms across successive retraining rounds. Round after round, the loop amplified popularity bias. "We then show how this bias amplification leads to several other problems such as declining the aggregate diversity, shifting the representation of users' taste over time and also homogenization of the users experience," the authors write.

The damage was not spread evenly. The effect was strongest for users in the minority group. So the arrow marked retraining carries three outcomes and an unequal incidence: aggregate diversity down, represented taste drifting, experiences converging — and the users furthest from the centre of the distribution absorbing most of it.

A head-only success metric hides the loop while it runs, and the reason is sharper than averages concealing detail. Common collaborative filters create a rich-get-richer effect that reduces aggregate sales diversity even while each individual user's diversity rises. Fleder and Hosanagar showed it analytically, then in simulation. "In line with this, we show it is possible for individual-level diversity to increase but aggregate diversity to decrease. Recommenders can push each person to new products, but they often push similar users toward the same products," they wrote, in Management Science in 2009.

So a rising per-user discovery dashboard beside a concentrating catalogue is not a paradox to be explained away in a review. It is the published signature of the failure. A team reading only the per-user number is watching the metric go up while the thing it stands for goes down. Intervention without follow-up hides the other half, because the diversity rule is withdrawn before the evidence about it has matured.

Draw the loop before measuring anything inside it. Step 1 maps the causal loop: exposure, behavior, training, provider and inventory pathways. Step 2 establishes distribution baselines — popularity, exposure, coverage, concentration and churn. Step 3 introduces controlled variation through exploration, cold-start allocation or policy rotation. Step 4 evaluates long horizons, covering user adaptation, provider response and catalog survival. Step 5 sets correction triggers, with declared thresholds for concentration, starvation and harm.

Steps 3 and 4 read as aspirations until somebody publishes both from inside a live product. A 2020 study did, inside Spotify. It covered over 100 million distinct Spotify users, who streamed around 70 billion times during the first 28 days of July 2019. Their listening was scored with a generalist-specialist diversity measure built from an embedding of over 850 million playlists. Algorithmically programmed listening was less diverse than organic listening for the vast majority of users, and the relationship held as intake rose. "Users who increase their intake of algorithmic recommendations become less diverse in their consumption," the authors write.

Then they ran step 3 for real. In a deployed randomised experiment, ranking by relevance rather than popularity raised song streams by 10.03% for generalists and 25.66% for specialists. The controlled variation did not cost engagement. It produced engagement — two and a half times as much of it for the users a popularity ranking serves worst. That is the shape of the answer an audit returns when exposure is varied on purpose. A number per population, not an argument about whether the loop exists.