Multi-Stakeholder and Marketplace Recommendation

An artificial music market put 14,341 participants in front of 48 songs by unknown bands. Participants in the social-influence condition were randomly assigned to one of eight worlds that then evolved independently. Same 48 songs in each. Same starting conditions. The only thing each world showed its own participants was what earlier participants in that world had chosen.

The eight worlds pulled apart. The same songs, offered to comparable people under identical initial conditions, ended up ranked differently in each one. The stronger the social signal, the wider the spread. Salganik and colleagues put it in one line in Science in 2006: “Increasing the strength of social influence increased both inequality and unpredictability of success.”

A marketplace ranking is that mechanism, running continuously, with money attached. It optimizes for purchases from mature listings. Established sellers collect more exposure and more reviews. New sellers cannot gather the evidence they would need to compete. The experiment is what tells you how much of the ordering is quality, and how much is the platform's own earlier output fed back in.

The system chooses not only what users see but which providers receive attention, demand, reviews and survival opportunities. Consumer utility, provider utility, platform value and ecosystem health can conflict. The size of that conflict has been measured.

Rank purely for the customer and the exposure collapses onto a few providers. FairRec simulated conventional top-k recommendation, k=20, on three settings. The data was Google Local: 11,172 customers, 855 Manhattan-area businesses, 25,686 reviews. And Last.fm: 1,892 customers, 17,632 artists, 92,834 play-count records. Patro and colleagues published what the exposure curves showed, in 2020: “We observe that the Lorenz curves for top-k recommendations are far below the equal exposure marks, revealing that for conventional top-k recommendation, 50% least exposed producers get only 32%, 5%, and 11% of total available exposure (m · k) in GL-CUSTOM, GL-FACT, and LF datasets, respectively.”

Half the producers sharing 5% of the available exposure is not a malfunction of that ranking. It is the ranking doing what it was asked to do.

Providers are not interchangeable either. A sale may be valuable to one seller and overload another. Capacity, quality, reliability, inventory and the long-term availability of alternatives all belong inside the recommendation.

The objective map holds five kinds of value at once: consumer, provider, platform, market structure, and the constraints and rights that bound all four. A marketplace decision that names only the first has not been made. It has been postponed.

The bottom layer is not a list of design preferences. Since 12 July 2020, the main parameters of a ranking have been a disclosable term of the provider relationship across the EU. The Platform-to-Business Regulation, Regulation (EU) 2019/1150, says it in Article 5(1): “Providers of online intermediation services shall set out in their terms and conditions the main parameters determining ranking and the reasons for the relative importance of those main parameters as opposed to other parameters.” Article 5(2) puts a parallel public-description duty on online search engines.

That turns the red layer from a value the team holds into an obligation the team owes.

Who counts as a stakeholder is settled first, and the list runs wider than buyers and sellers. It reaches workers, the platform itself, and the markets around it. Monitoring at the far end watches concentration, entry, exit, regional coverage and complaints.

Step 4 — transparent limits, quotas or re-ranking where justified — has a shipped instance with numbers attached. LinkedIn built fairness-aware re-ranking into Talent Search and published the results in 2019. Online A/B tests measured it first. Then it went to every LinkedIn Recruiter user worldwide, across a member base of more than 630 million. Geyik and colleagues state the outcome in the abstract: “Our approach resulted in tremendous improvement in the fairness metrics (nearly three fold increase in the number of search queries with representative results) without affecting the business metrics, which paved the way for deployment to 100% of LinkedIn Recruiter users worldwide.”

Step 4 does not promise that re-ranking is free. It demands that the consumer-side cost be measured. Here the measurement came back at none detected.

Report user utility, provider exposure, provider outcomes, capacity-adjusted value, concentration and new-provider survival. Slice by market, category, seller size and inventory state.

Estimating long-term supply effects may need randomized or phased policy changes. A short A/B test cannot fully capture whether providers invest, enter, leave or change behavior.

The horizon problem reaches the definition of fairness itself. No single ranking can achieve individual attention fairness, Biega and colleagues argued in 2018, so they amortized attention across a series of rankings instead. Their reason: “As ranking positions influence the amount of attention the ranked subjects receive, biases in rankings can lead to unfair distribution of opportunities and resources such as jobs or income.” They tested on real Airbnb supply: 3,944 listings in Boston, 1,728 in Geneva and 4,529 in Hong Kong. Each was ranked on seven separate review-rating attributes — review_scores_rating, accuracy, cleanliness, checkin, communication, location and value.

On marketplace supply of that size, the accounting unit that can carry a fairness claim is the sequence of slates, not the impression.

Being seen and being chosen are different things to owe a provider. Exposure parity or allocation controls visible opportunity, and can be measured on the slate itself. Outcome support targets sales, bookings, applications or qualified leads. Those depend on how consumers respond and on what the provider can actually deliver. Marketplace health is the slowest of the three, and cannot be attributed to any single slate.

Exposure is worth governing because position weighting is not a gentle transformation of merit. Singh and Joachims open a 2018 paper with six job applicants. Their relevances are 0.80, 0.79, 0.78 and 0.77, 0.76, 0.75. Apply a standard 1/log(1+j) position-bias drop-off and the lower-scored group receives roughly 30% less exposure. A 0.03 gap in average relevance has become a 0.32 gap in average exposure. An order of magnitude of opportunity, conjured from a rounding error of merit.

Their question about that ranking is the one a marketplace has to answer for every slate it prints: “Is this winner-take-all allocation of exposure fair in this context, even if the winner just has a tiny advantage in relevance?”