← Paper page How Bad is Top-K Recommendation under Competing Content Creators? Ask this paper arXiv ↗

How Bad is Top-K Recommendation
under Competing Content Creators?

When creators chase attention, do users still get good recommendations?

Fan Yao, Chuanhao Li, Denis Nekipelov, Hongning Wang, Haifeng Xu · ICML 2023 (oral)

Press → or click to step through

Three players on a recommendation platform

creators choose what to make platform ranks content by relevance top K = 3 users pick one, with some randomness creators are paid by how much users engage

Everyone acts in their own interest. What happens to users?

The worry: everyone makes the same thing

What users need 4 sports fans 2 travel fans sportssportssports travel every user finds something good ✓
What competition can produce sportssportssports sports travel fans left out the majority topic crowds out the niche

How much user welfare can this kind of competition destroy?

Earlier models were pessimistic: the loss could be unbounded, or up to half.

Result: at most half is ever lost, and usually far less

worst-case share of the best possible user welfare that competition keeps best possible K = 1 ≥ 50% K = 2 ≥ 56% K = 3 ≥ 61% K = 4 ≥ 64% K = 5 ≥ 66% K = 7 ≥ 69% guarantees proved in the paper, with users’ choice randomness β = 0.5

Showing users more options (larger K) protects them.

The guarantee also improves when users’ choices are less predictable; the worst case needs K = 1 or users who always pick the top item.

The bound is tight: the paper builds examples that come close to it.

It holds even while creators are still learning

time → what each creator makes average user welfare stays within the guarantee

Creators don’t know their payoffs in advance; they learn by trial and error.

The guarantee covers any sensible learning (“no-regret”) and needs no stable equilibrium. Schematic, not data.

What creators are paid for matters

paid per view exposure loss can exceed half
paid per satisfied view engagement stays within the guarantee

In simulations, paying for exposure lost more than half the welfare once there were over 10 creators; paying for engagement did not.

The gap is largest when popular content is cheap to make.

On MovieLens data, the loss is small

welfare lost to competition, relative to the best possible top-5 recommendation, β = 0.1; creators learn with an exploration rate; averages of 10 runs (paper, Table 5)
creators’ exploration rate

Even with five creators the loss is about 10%, and it shrinks as more creators compete.

Exploring too little or too much makes things worse: try the buttons.

What this means for platforms

  1. 1Plain relevance ranking holds up well in the long runSelfish creators never cost users more than half, and usually much less.
  2. 2Show several optionsDiversity in the list, together with users’ own randomness, is what protects welfare.
  3. 3Reward engagement, not exposureThe incentive design decides whether competition helps or hurts.

Paper page · arXiv · ICML talk