Dynamics in Two-Sided Attention Markets: Objective, Optimization, and Control

Haiqing Zhu, Yun Kuen Cheung, Lexing Xie · 2025

Read Dynamics in Two-Sided Attention Markets: Objective, Optimization, and Control by Haiqing Zhu, Yun Kuen Cheung, Lexing Xie (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-01970.txt.

Introduction

Haiqing Zhu, Yun Kuen Cheung, and Lexing Xie model two-sided attention markets where platforms mediate users and creators. Most content now flows through such platforms, yet theory often covers only one side at a time. Past work mapped creator or user interactions alone, with static balances and learning rules. Their model couples all three parties in one system.

A potential function captures the coupled play among users, platform ranking, and creators. Creator best response moves paired with user choice rules turn out to equal mirror descent on that potential. Whole families of ranking rules fit the same frame with changed potentials. A new local convergence result for mirror descent on nonconvex functions supports the analysis. The frame helps explain the varied outcomes seen in real attention markets.

Catalog record

CREATORHaiqing Zhu, Yun Kuen Cheung, Lexing Xie
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE96 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-01970

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamics in Two-Sided Attention Markets: Objective, Optimization, and Control Haiqing Zhu, Yun Kuen Cheung, Lexing Xie Abstract With most content distributed online and mediated by platforms, there is a pressing need to understand the ecosystem of content creation and consumption. A considerable body of recent work shed light on the one-sided market on creator-platform or user-platform interactions, showing key properties of static (Nash) equilibria and online learning. In this work, we examine the two-sided market including the platform and both users and creators. We design a potential function for the coupled interactions among users, platform and creators. We show that such coupling of creators’ best-response dynamics with users’ multilogit choices is equivalent to mirror descent on this potential function. Furthermore, a range of platform ranking strategies correspond to a family of potential functions, and the dynamics of two-sided interactions still correspond to mirror descent. We also provide new local convergence result for mirror descent in non-convex functions, which could be of independent interest. Our results provide a theoretical foundation for explaining the diverse outcomes observed in attention markets. 1 Introduction We are concerned with attention markets of online content – where the key resource being allocated is users’ attention (and time). This market is two-sided . On one-side, a platform distributes content to users, controlling content visibility via recommender systems; on the other side, the platform syndicates content from creators and distributes reward back to them. Feedback loops tie the two sides together. Our main concern and contribution lies in an optimization view of such two-sided interactions: is there a global objective, do myopic updates minimise any objective, and do different recommendation strategies change the overall optimisation problem? The answers to all three question are affirmative, but we shall first relate our questions to what has been asked about markets. In attention markets , we focus on the allocation of users’ time and creators’ effort, neither of which is quantified in monetary terms but is the key driver to most monetary incentives. This is distinct from the large body of work on advertising-driven attention Varian 2009 , or charging schemes on either side Vogelsang 2010 . In economics, markets in which platforms act as intermediaries between two groups of agents are known as two-sided markets 1 1 1 These markets are different from the “two-sided markets” in bipartite matching context. . Our concern in two-sided markets is a special case with one platform mediating a large number of players on either side, and hence distinct from known literature on platform competition Rochet and Tirole 2003 ; Armstrong 2006 . Furthermore, price is not the key mediator in our market, but bounded attention is. We assume that the total attention available is bounded rather than infinite Ghosh and Hummel 2014 , which is arguably closer to the original notion of attention markets Simon 1971 . In comparison, there is unbounded supply of each digital good, assuming that each can be copied and consumed by many people at negligible additional cost. Figure 1 contains a high-level overview of the two-sided attention market (detail in Section 2 ). We take an optimization perspective to examine the intertwined dynamics of user-platform and creator-platform interactions. In our model, users respond to the visibility , quality and popularity of items. Aggregating user choices yields popularity signal across all items. Creators are rewarded proportional to the popularity of items they create, and respond with a choice in item quality sensitive to cost of production. Platforms control item visibility , while relaying quality and popularity signals to both users and creators. Prior research largely focused on one side of the market, such as user-platform side Zhu et al. 2023 or creator-platform side Ben-Porat and Tennenholtz 2018 ; Yao et al. 2023b ; Jagadeesan et al. 2023 . Our user model follows a trial-offer market dynamic Krumme et al. 2012 inspired by the Music Lab experiment Salganik et al. 2006 . Content quality in this work assumes a scalar value; we leave vector-valued quality, such as those produced by personalisation and recommender systems McAuley 2022 , as future work. Our creator reward vs cost model is inspired by Ghosh and Hummel 2014 except assuming bounded total reward. It affords the most general form of production cost, rather than assuming a particular functional form such as Jagadeesan et al. 2023 ; Yao et al. 2023a . Game-theoretic analysis Ben-Porat and Tennenholtz 2018 ; Yao et al. 2023a have provided useful insights including the qualities of equilibria and their convergence dynamics under no-regret learning. But our quest for an overall optimization problem provides additional benefits on predicting the long-term dynamics, which in turn could inform the design of effective policy interventions that promote a healthy ecosystem, reduce the spread of undesirable content, and maintain fair income distributions among creators. Figure 1: An overview of two-sided attention market. Users chose content based on content visibility and quality to produce popularity signals. The platform updates visibility and pass along popularity to creators, which in turn drives creators to update content quality . Notations correspond to those in Section 2 . Our first result is uncovering a potential function underlying the two-sided attention market ( Equation 16 ), which is a combination of expected log-utility, cost of production, entropy, and the alignment between attention and visibility. This potential function is similar in spirit to those found in a potential game in that it expresses a global incentive for all parties to take action. We identify a transformation to express the function in a lower-dimensional space of user actions, rather than the joint action space of all players ( Section 3.2 ). The second result shows that user and creator update dynamics correspond to mirror descent, a fundamental optimization algorithm, on the potential function ( Theorem 16 ). Finally, we show that when the platform vary its ranking strategies combining quality and popularity, there is a family of potential functions on which mirror descent (with momentum) still describes the corresponding response dynamic by users and creators ( Section 3.4 ). To understand market dynamics governed by the non-convex potential functions, we complement the existing mirror descent toolbox by providing local convergence result ( Theorem 14 ), which may be of independent interest. Figure 2: Illustration of three instances of two-sided market dynamics. Each triangle represents the projection of the probability simplex of a 3-item market in \mathbb{R}^{3} . The contours within each simplex represent the values of the potential function of that market, where darker colours indicate smaller function values. The trajectories demonstrate the two-sided market dynamics with three different initialisations, where “ \bullet ” mark the initial points. The potential function is non-convex in general, meaning the market dynamics can reach different local minimisers, which correspond to equilibria with different properties. Figure 2 demonstrates the market dynamics under three instances of the markets. The contour plots represent the value of the corresponding potential functions defined in Equation 15 . From left to right, the social influence factors r=0.1,0.38 and 0.5 . Similar landscapes of potential functions could also be equivalently generated by changing the costs of production \mathbf{c} or implementing appropriate recommendation policies (cf. Section 2.4 ). As shown in Figure 2 , though

Read the complete text