Seyed A. Esmaeili

I'm a Postdoctoral Researcher at the Data Science Institute at the University of Chicago where I am hosted by Haifeng Xu.

Previously I was a Postdoctoral Fellow at the Simons Laufer Mathematical Sciences Institute (SLMath) at Berkeley where I was affiliated with the Algorithms, Fairness, and Equity program.

Before that I was a CS PhD student at the University of Maryland where I worked with John P. Dickerson and Aravind Srinivasan.

Email: last_name@uchicago.edu

Research

I'm interested in machine learning, algorithmic fairness, and strategic and game theoretic aspects of machine learning. More specifically, I have worked on fair clustering, redistricting/gerrymandering, and game-theoretic/strategic issues in multi-armed bandits. Recently, I've been working on applying large language models in game-theoretic settings as well as studying their social impacts.

We study a variant of the multi-armed bandits problem where the arms are strategic and have control over their output level. We give algorithms that achieve the three objectives of learning, performance incentivization, and robustness simultaneously.

We study the canonical fair clustering problem when the group memberships are not perfectly known. Specifically, we introduce an interpretable family of error models that require little input from the decision maker and use robust optimization to produce algorithms with theoretical guarantees.

We study a bandit setting where agents can possibly register an unbounded number of arms. The agents have Bayesian priors over the arms' quality levels (means) but do not see the realizations of the means before registration. We show that prior work fails in this setting and give algorithms that are replication-proof.

We identify a collection of shortcomings in the fair clustering literature and show a set of examples where the application of a fair clustering algorithm could possibly even degrade the welfare. Finally, we give some recommendations that could lead to more impactful fair clustering research.

The fair clustering literature has produced a large number of fairness notions and has largely ignored how these notions might be satisfied simultaneously or how they relate to one another. In this project, we choose two specific group representation-based fairness notions and give algorithms that can satisfy both of them simultaneously. In addition, we show incompatibility between these two notions and a set of distance-based fairness notions.

We introduce a simple and interpretable distance measure over redistricting maps. This enables us to select a central map that mirrors the Kemeny ranking in a scenario where a committee votes over a collection of maps. A byproduct of our framework is the possible detection of gerrymandered maps as we show that some instances are outliers in terms of distance. Our framework has significant computational speedups in comparison to prior work.

We consider an online matching platform and show algorithms that give fairness guarantees for the two sides to be matched as well as a revenue guarantee for the platform operator. We further establish lower bounds on the performance of any algorithm and show that in general individual and group fairness can be at odds with one another.

We introduce a new notion of individual fairness in clustering which constraints the solution to have equitable distance to center values across the points. We establish lower bounds on our new clustering problem, study its price of fairness, and give approximation algorithms for it.

We consider a setting where different clusters have outcomes of possibly varying qualities (labels) associated with them. We show that unlike previous work, fairness considerations are more affectively satisfied by having close to population level proportional representation of each group across the label not the clusters. We further give algorithms with theoretical guarantees for this setting.

Imposing a fairness constraint on clustering may lead to an unbounded degradation in the clustering cost. Therefore, in this project we introduce a new problem where we maximize the amount of fairness subject to an upper bound on the clustering cost. We define appropriate fairness measures and give algorithms with guarantees for them.

We study the fair clustering problem when the group memberships are imperfectly known. Specifically, we consider the fairness notion where each cluster is supposed to have close to population level proportion of each group in the setting where probabilistic instead of deterministic group membership knowledge is available. We introduce a probabilistic fairness notion for this setting and give algorithms with theoretical guarantees to solve it.

We introduce a differentially private matrix/tensor-based spectral algorithm for Latent Dirichlet Allocation. We present the algorithm as a computational graph and study noise injection at different edges establishing utility guarantees for each different choice and showing that some can outperform others at different regimes.

We solve the problem of automated thumbnail generation using a deep neural net achieving higher quality thumbnails with a faster run-time than previous method. We also introduce a dataset of images and their manually generated thumbnails.

Website template from Jon Barron's website.