Mihaela Curmei

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-159

August 7, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-159.pdf

Algorithmic decision systems are an ubiquitous part of daily life, operating in dynamic environments where user interactions and feedback loops complicate predictability, reliability, and utility. This thesis explores the multifaceted dynamics between users and algorithmic decision systems, addressing both immediate impacts and long-term implications of their interactions. By examining single-user interactions and extending to broader social network and market dynamics, we aim to uncover trade-offs and limitations in these feedback settings. This comprehensive study provides insights into these interactions, emphasizing the importance of considering both individual and collective behaviors in system design and evaluation.

Part I focuses on user-recommender interactions, providing insights into how users and recommendation systems affect each other. We introduce a user-centric notion of agency in algorithmic recommendations, focusing on the feasible outcomes of one-step interactions between users and systems. By proposing an evaluation procedure based on stochastic reachability, we quantify the maximum probability of recommending a target piece of content to a user under allowable strategic modifications. This framework allows us to detect biases and systemic limitations in content discovery with minimal assumptions about user behavior. Transitioning from single-step interactions, we explore how recommendations influence users in multi-step closed-loop dynamics, requiring modeling of user behavior. We develop psychologically grounded dynamic preference models to capture classic psychological effects such as Mere Exposure, Operant Conditioning, and Hedonic Adaptation. Simulation-based studies show these models manifest distinct behaviors, informing system design and allowing for critical evaluations based on psychological plausibility.

Part II broadens the scope to consider the interactions of multiple users and multiple algorithmic decision-makers, examining complex social network and market dynamics. First, we investigate social networks where user interactions are mediated by link recommendation algorithms, examining the interplay of multiple users and relationships within evolving networks. Using an extended Jackson-Rogers model, we evaluate how link recommendations influence network evolution over time, revealing delayed and indirect effects on network structures. Additionally, we examine markets with multiple algorithmic decision-makers, analyzing dynamics where users allocate their participation among services to minimize individual risk while services update their models to reduce risk based on current user populations. Termed risk-reducing dynamics, this class includes common model updates such as gradient descent and multiplicative weights. Our findings indicate that repeated myopic updates with multiple learners result in market segmentation as the only stable outcome. We argue that specialization is an emergent property of competition, which alleviates typical concerns of representation disparity.

Advisors: Benjamin Recht

---

BibTeX citation:

@phdthesis{Curmei:EECS-2024-159,
    Author= {Curmei, Mihaela},
    Title= {Feedback Driven Dynamics in Socio-Algorithmic Systems},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-159.html},
    Number= {UCB/EECS-2024-159},
    Abstract= {Algorithmic decision systems are an ubiquitous part of daily life, operating in dynamic environments where user interactions and feedback loops complicate predictability, reliability, and utility. This thesis explores the multifaceted dynamics between users and algorithmic decision systems, addressing both immediate impacts and long-term implications of their interactions. By examining single-user interactions and extending to broader social network and market dynamics, we aim to uncover trade-offs and limitations in these feedback settings. This comprehensive study provides insights into these interactions, emphasizing the importance of considering both individual and collective behaviors in system design and evaluation.

Part I focuses on user-recommender interactions, providing insights into how users and recommendation systems affect each other. We introduce a user-centric notion of agency in algorithmic recommendations, focusing on the feasible outcomes of one-step interactions between users and systems. By proposing an evaluation procedure based on stochastic reachability, we quantify the maximum probability of recommending a target piece of content to a user under allowable strategic modifications. This framework allows us to detect biases and systemic limitations in content discovery with minimal assumptions about user behavior. Transitioning from single-step interactions, we explore how recommendations influence users in multi-step closed-loop dynamics, requiring modeling of user behavior. We develop psychologically grounded dynamic preference models to capture classic psychological effects such as Mere Exposure, Operant Conditioning, and Hedonic Adaptation. Simulation-based studies show these models manifest distinct behaviors, informing system design and allowing for critical evaluations based on psychological plausibility.

Part II broadens the scope to consider the interactions of multiple users and multiple algorithmic decision-makers, examining complex social network and market dynamics. First, we investigate social networks where user interactions are mediated by link recommendation algorithms, examining the interplay of multiple users and relationships within evolving networks. Using an extended Jackson-Rogers model, we evaluate how link recommendations influence network evolution over time, revealing delayed and indirect effects on network structures. Additionally, we examine markets with multiple algorithmic decision-makers, analyzing dynamics where users allocate their participation among services to minimize individual risk while services update their models to reduce risk based on current user populations. Termed risk-reducing dynamics, this class includes common model updates such as gradient descent and multiplicative weights. Our findings indicate that repeated myopic updates with multiple learners result in market segmentation as the only stable outcome. We argue that specialization is an emergent property of competition, which alleviates typical concerns of representation disparity.},
}

EndNote citation:

%0 Thesis
%A Curmei, Mihaela 
%T Feedback Driven Dynamics in Socio-Algorithmic Systems
%I EECS Department, University of California, Berkeley
%D 2024
%8 August 7
%@ UCB/EECS-2024-159
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-159.html
%F Curmei:EECS-2024-159