← Paper page Policy Design for Two-sided Platforms with Participation Dynamics Ask this paper arXiv ↗

Policy Design for Two-sided Platforms
with Participation Dynamics

When viewers and creators come and go depending on what the platform recommends, how should it recommend?

Haruka Kiyohara, Fan Yao, Sarah Dean · ICML 2025

Press → or click to step through

A platform between two crowds

viewers in groups, e.g. sports fans providers e.g. sports creators platform its policy decides which providers each viewer group gets to see viewers get satisfaction providers get exposure

Viewers want good content. Providers want to be seen.

Examples: video streaming, e-commerce, job matching. The platform’s goal is total viewer satisfaction (its social welfare).

Both crowds grow or shrink, and feed each other

more viewers join providers get more exposure more providers join better picks for viewers population effects a self-reinforcing loop

A bigger pool of providers means better content for viewers, and more viewers mean more exposure for providers.

It also runs in reverse: providers with little exposure leave, and viewers get worse picks.

Earlier models kept one side fixed. Here both move, and the paper proves the dynamics settle at a stable point when populations adjust gradually.

The default: be greedy for today

Myopic-greedy everyone sees today’s top group no exposure: those providers leave A B C
Spreading exposure each group gets some attention A B C every group can grow

Greedy matches each viewer group with what looks best right now, so exposure piles up on a few provider groups.

Spreading exposure keeps several provider groups growing, and viewers can gain from the bigger pool later.

Schematic. The paper shows that concentrated exposure can polarize the platform into winners and losers and shrink the pie.

Greedy is safe only in a special case

Every group improves alike quality for viewers providers in the group → A B then greedy is (near-)optimal
Groups improve differently quality for viewers providers in the group → A B B looks better today A, once grown greedy feeds B; A never grows

Greedy counts the loss from today’s choice, but ignores the loss from growing the wrong crowd.

Schematics of the paper’s Theorem 3 (linear, identical gains), Proposition 2 (differing gains) and Theorem 2 (the shortfall splits into a “policy” part and a “population” part).

The fix: judge a policy by the crowd it will grow

today’s populations try a policy satisfaction & exposure where the crowds are heading the key step welfare there keep the policy that scores best today only (greedy) long run (look-ahead) a dial blends the two

Score each candidate policy by the populations it is steering toward, not just today’s.

The platform can blend look-ahead with greedy; pure look-ahead worked well in the experiments.

If viewers’ and providers’ responses are unknown, the platform first explores, learns them from data, then commits.

In simulation, looking ahead grows the pie

viewers providers one icon ≈ 100 providers on the platform 05001,0001,500 total viewer welfare 20k30k40k50k 220200 220200 time steps (log scale)time steps (log scale)

Greedy halves the provider side; look-ahead grows both sides and ends with about twice the welfare.

The paper’s synthetic test (20 viewer and 20 provider groups, small starting populations); values read off its Figures 1 and 2, approximate. Uniform ends with slightly more providers, but look-ahead spends exposure more efficiently and earns more welfare.

On real viewing data: high welfare and a healthy provider side

total viewer welfare at the end greedy ≈ 9,400 look-ahead ≈ 9,200 uniform ≈ 4,500 providers on the platform at the end greedy ≈ 800 look-ahead ≈ 1,800 uniform ≈ 1,850 approximate end-of-run values read off the paper’s Figure 5 (KuaiRec data)

Here greedy beats uniform on welfare. Look-ahead matches greedy and keeps almost as many providers as uniform.

KuaiRec: 4,676,570 interactions between 1,411 viewers and 3,326 videos, grouped into 20 viewer and 20 provider groups; the platform had to learn how populations respond. A 60/40 blend of look-ahead and greedy did best overall, and every blend did well.

What this means for platforms

  1. 1Recommendations shape who shows upViewers and providers grow or shrink with the satisfaction and exposure they get.
  2. 2Greedy can shrink the pieIt is only safe when every provider group gains from growth in the same steady way; concentrated exposure can drive providers away and hurt viewers.
  3. 3Look ahead, and spread exposure wiselyScoring policies by the populations they will create grew both sides, and welfare, in the experiments.

Paper page · arXiv · Code · Slides · Interview