Multistakeholder Impacts of Profile Portability in a Recommender Ecosystem Conference: UMAP’26 Conference; June, 2026; Gothenburg CCS: Information systems Recommender systems CCS: Computing methodologies Multi-agent systems CCS: Social and professional topics User characteristics

Anas Buhayh, Elizabeth McKinnie, Clement Canel, Robin Burke · 2026

Read Multistakeholder Impacts of Profile Portability in a Recommender Ecosystem Conference: UMAP’26 Conference; June, 2026; Gothenburg CCS: Information systems Recommender systems CCS: Computing methodologies Multi-agent systems CCS: Social and professional topics User characteristics by Anas Buhayh, Elizabeth McKinnie, Clement Canel, Robin Burke (2026) 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-2604-21750.txt.

Introduction

Buhayh and three colleagues examine what happens when recommender users can take their profile with them and pick their own algorithm apart from the platform. Earlier work suggests such choice helps niche consumers and providers. The paper asks what happens to user models when people switch algorithms and their data moves too. The effects on user benefit differ across algorithms, raising policy questions for fair recommender design.

Catalog record

CREATORAnas Buhayh, Elizabeth McKinnie, Clement Canel, Robin Burke
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46173
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.21750

Browse the full library · About the Tower of Babel

Full text

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

Multistakeholder Impacts of Profile Portability in a Recommender Ecosystem Conference: UMAP’26 Conference; June, 2026; Gothenburg CCS: Information systems Recommender systems CCS: Computing methodologies Multi-agent systems CCS: Social and professional topics User characteristics Anas Buhayh, Elizabeth McKinnie, Clement Canel, Robin Burke Abstract . Optimizing outcomes for multiple stakeholders in recommender systems has historically focused on algorithmic interventions, such as developing multi-objective models or re-ranking results from existing algorithms. However, structural changes to the recommendation ecosystem itself remain understudied. This paper explores the implications of algorithmic pluralism (also known as "middleware" in the governance literature), in which recommendation algorithms are decoupled from platforms, enabling users to select their preferred algorithm. Prior simulation work demonstrates that algorithmic choice benefits niche consumers and providers. Yet this approach raises critical questions about user modeling in the context of data portability: when users switch algorithms, what happens to their data? Noting that multiple data portability regulations have emerged to strengthen user data ownership and control. We examine how such policies affect user models and stakeholders’ outcomes in recommendation setting. Our findings reveal that data portability scenarios produce varying effects on user utility across different recommendation algorithms. We highlight key policy considerations and implications for designing equitable recommendation ecosystems. 1. Introduction There is substantial literature on multistakeholder recommender systems, outlined in ( 1 ) . Various interventions have been designed to balance system benefits across stakeholders. We see two main types of interventions: model-based (incorporating multistakeholder objectives into the recommendation model) and re-ranking (applying post-hoc adjustments to recommendation lists). An alternative point of intervention is at the ecosystem level: the multistakeholder properties of an individual recommender system may be less critical if a variety of recommendation algorithms with different properties are available. Market-based recommendation ecosystems have been proposed under various names: friendly neighborhood algorithm store ( 46 ) , middleware ( 23 ; 30 ) , or algorithmic pluralism ( 57 ) . As in other economic sectors, we might expect algorithm designers to specialize in serving particular audiences, and stakeholder outcomes — both consumer-side and provider-side — may be enhanced over solutions that rely on a single recommender system to be all things to all people. Using a simulation-based approach, 12 ; 13 explored the dynamics of a recommender system marketplace with two algorithm providers: one for a generic audience and one focused on a particular content niche. These studies found that a multistakeholder recommender ecosystem created fairer outcomes for users with distinct tastes outside of the mainstream, and those who produced content catering to niche tastes were able to achieve better outcomes with minimal loss to the utility experienced by other consumers and providers. In this work, we extend this line of research by investigating how regulatory choices, specifically around user profile management, impact a multi-algorithm recommender ecosystem and the outcomes for niche consumers and providers. In particular, we study the question of profile portability: to what extent are user profiles shared (or not) among algorithm operators when users move from one algorithm to another? We explore two dimensions of portability. The first is exclusivity , the extent to which a user’s profile data remains tied to a given algorithm and does not follow the user when switching to a new recommender. Since algorithm operators compete, they may be unwilling to share user data with rivals, creating significant ‘lock-in’ as users may hesitate to switch if their data cannot move with them. Article 20 of the GDPR ( 44 ) grants users the right to download or transfer their data, which may be interpreted as requiring non-exclusivity. The second aspect of portability is permanence : whether user data is retained by a system when a user switches algorithms. Data accumulation is crucial for large-scale machine learning platforms such as recommender systems, and operators are likely to retain as much data as possible, even after users leave. However, countervailing principles exist: Article 5(1)(c) of the GDPR establishes a data minimization principle, requiring data to be limited to what is necessary, and Article 5(1)(e) restricts retention to no longer than necessary. It could be argued that a departing user’s profile is no longer needed by the original platform, implying non-permanence. The interaction between exclusivity and permanence gives us four different conditions as depicted in Figure 1 . 1 1 1 Icons courtesy of The Noun Project. Figure 1 . Profile portability options. Diagram titled “Profile portability options” showing four scenarios in a 2×2 grid. Top row labeled “Exclusive? Yes,” bottom row “No.” Columns show different portability designs: Algorithm-specific (top-left), Cold start (top-right), Universal (bottom-left), and User ownership (bottom-right). \parEach panel includes a user icon, two boxes labeled “Algorithm A” and “Algorithm B,” and one or more “Profile” shapes connected by arrows. In Algorithm-specific, each algorithm has its own separate profile. In Cold start, switching breaks the connection to the existing profile (marked with a red X) and creates a new empty profile. In Universal, both algorithms share a single common profile. In User ownership, the profile moves with the user from one algorithm to another, with the previous connection removed (red X) and a new one established. Algorithm-Specific (Exclusive/Permanent): This reflects a competitive, unregulated market where users lack control over their profiles. Algorithms retain data when users switch and do not allow it to be transferred to competitors. Cold Start (Exclusive/Non-permanent): This is the most restrictive case: users cannot transfer data, and profiles are deleted upon switching, making every user a cold-start case. User Ownership (Non-exclusive/Non-permanent): Users carry their profiles across algorithms, creating blended histories, while platforms cannot access data of non-subscribers. Users effectively own and transfer their data. Universal Profile (Non-exclusive/Permanent): Users bring their data, but platforms retain access even after users leave, resembling a shared profile database. This may arise in shared infrastructures ( 30 ) . These designs are not hypothetical. Algorithm-Specific is standard, with platforms owning training data. Cold Start appears in systems that allow data deletion, studied as machine unlearning ( 10 ) . Universal designs occur in shared-protocol systems or where multiple algorithms operate on one platform. User Ownership exists in ecosystems enabling cross-platform data transfer, such as importing music into Spotify via third-party tools 2 2 2 https://www.spotify.com/us/import-music/ These conditions give rise to research questions that we explore in this work. Following the example of ( 12 ) , we concentrate on the simplified case where there are only two recommenders: one serving the general population – we call this the Generic recommender R^{G} – and a second Niche recommender system R^{N} that has been created to meet the needs of a subset of users for whom the generic one does not serve well. RQ1 : How does the exclusivity of a profile impact outcomes for consumers transitioning from one recommender to another? We might expect that having a profile tied to a particular recommender would mean that consumers who switch will receive lower-quality recommendations. RQ2 : How does the permanence of a profile impact consumers who are not changing recommenders? If profiles are not permanent, consumers (e.g., Algorithm A consumers in the User Ownership case) receive recommendations generated from a smaller, potentially less diverse dataset after Niche consumers have left. RQ3 : How do these profile options impact the

Read the complete text