My research philosophy is to work on problems that not only advance our understanding of the fundamental questions in optimization and machine learning, but often also has a positive impact on the society. My work tackles important theory questions, that are often motivated by practice. The two broad research areas that I work in are:

A. Robust, Reliable and Trustworthy Algorithms,

B. Efficient Optimization and Learning.

In the former, I aim to mitigate the impact of uncertainty and noise in automated decision-making systems that can lead to unintended consequences, using models and algorithms that balance efficiency with robustness, reliability, and trustworthiness. In the latter, I develop fast algorithms tailored to large-scale, evolving societal challenges by leveraging the combinatorial structure of problems. I next discuss some research directions we are actively pursuing in my group. (For a list of publications, visit Publications.)  

A. Robust, Reliable and Trustworthy Algorithms

Two key directions that we are interested in so far are:

Theme 1. Mitigating the Impact of Noisy and Unreliable Data, and
Theme 2. Multi-Criteria Decision-Making Competing Objectives.

To tackle noise and uncertainty in observed data, we develop new models to capture variability in data due to various socio-economic reasons and analyze these to understand the impact on algorithmic decisions (e.g., stable matchings of students to schools) for different groups of people. This analysis is key in downstream decisions, e.g., in planning investment of limited resources. I received the NSF CAREER Award in 2023 to explore these directions further. Next, my work studies the trade-offs between competing goals like optimizing utilitarian criteria (e.g., cost minimization) and equity-inducing properties such as monotonicity, balance, envy-freeness. This has led to new results on approximability of discrete problems with small portfolios of solutions, on algorithms for trajectory-constrained online optimization, and tractable methods for districting. Given the socio-economic nature of FATE problems, we collaborate with law and public policy researchers as well, which I will discuss below.

Theme 1. Mitigating the Impact of Biased and Noisy Data: The key question we are interested in is how to make “robust and reliable” decisions when the observed data is noisy or biased. Evaluation data often shows distributional differences across individuals from different backgrounds. However, algorithmic techniques to make data-driven decisions on such data, without exacerbating unfairness are not well-understood. This research theme explores new algorithms under different distributional assumptions, characterization of impact of presence of noise in data, and interventions for reducing these impacts. The problems with noisy data impact downstream decision-making pipelines, often involving some machine learning. Some times domain-specific constraints or high-quality data can be used to augment noisy pipelines for more resilient and robust decisions. In this theme, we explore the trade-offs between quality of data and the cost of acquiring it through robust and resilient algorithms. Here are some data models we have considered thus far:

(a) Ordinal Online Optimization: Currently, 97% organizations are rely on automated algorithms for candidate selection process, as it is impossible for humans to sift through millions of resumes or test scores or health records. Unfortunately, predicting “hirability” for each candidate creates multiple challenges. These ML generated scores tend to pick up biased trends in the underlying data. This poses a huge challenge as a significant fraction of the U.S. workforce, with at least 70+ million adults in the U.S., are “STARS” skilled through alternative routes, such as community college, workforce training, boot camps, certificate programs, military service, or on-the-job learning, rather than through a traditional 4-year bachelor’s degree. Proactively addressing these challenges remains a big open question.

From a theoretical lens, we are broadly interested in extending existing algorithms for ordinal data (i.e., data where only comparisons can be performed, without numeric values). We explored this in the context of hiring, specifically for the online secretary problem. As opposed to existing literature, where the algorithm can observe utility of candidates, we restricted to the setting where the algorithm can only observe a partial order over the candidates that have arrived thus far. We showed that competitive algorithms in this setting are naturally fairer (the candidates higher up in the poset have a higher probability of selection), induce better data-specific distributions, and do not overcorrect the selection rates across gender groups (as opposed to using quotas). We further showed legal implications of this work.

poset

  • Secretary Problems with Biased Evaluations using Partial Ordinal Information, Jad Salem, Swati Gupta. Management Science, 2023.
  • Closing the GAP: Mitigating Bias in Online Resume-Filtering, Jad Salem, Swati Gupta. WINE 2020. link
  • Using Algorithms to Tame Discrimination: A Path to Algorithmic Diversity, Equity, and Inclusion – Deven Desai, Swati Gupta, Jad Salem, UC Davis Law Review, 2023
  • Don’t let Ricci v. DeStefano Hold You Back: A Bias-Aware Legal Solution to the Hiring Paradox – Jad Salem, Deven Desai, Swati Gupta, FAccT 2022

(b) Impact and Interventions for Noisy Data: We are interested in estimating bias in data, and developing interventions to mitigate its impact. Using data from the department of education in New York, we showed that there are consistent distributional differences in student performance on the specialized high school SAT (SHSAT), based on whether the students went to schools with low economic need (G1) or high economic need (G2). Our recent work explores the question about which students are the most impacted if this distributional difference persist and how can one use scholarships or additional trainings to mitigate these impacts. This involves understanding the stable matching mechanism under noisy student performance. We further explored the matching mechanism of the discovery program, and showed that this can in fact creating a large number of blocking pairs within the group of disadvantaged students (a blocking pair is when a student and school would rather be matched to each other compared to their current matches). We characterized new market conditions of “high competitiveness” under which another mechanism would provably not create such blocking pairs.

We are further exploring bias in organ transplantation pipelines and potential interventions, in collaboration with MGH.

  • Discovering Opportunities in New York City’s Discovery Program: an Analysis of Affirmative Action Mechanisms, Yuri Faenza, Swati Gupta and Xuan Zhang, EC 2023 (journal version under submission). arxiv
  • Reducing the Filtering Effect in Public School Admissions: A Bias-aware Analysis for Targeted Interventions, Yuri Faenza, Swati Gupta, Aapeli Vuorinen, Xuan Zhang, ACDA 2023 (journal version under revision at M&SOM) arxiv

(c) Incorporating Domain Constraints: Errors in measurements, like pulse oximetry errors for darker-skinned individuals, can lead to unintentional biases in patient care, and electronic medical records data is rife with such errors and recorded noise. This is a massive challenge today as machine learning pipelines are permeating the space of life-critical decisions. In a collaboration with the Emory University and Grady Hospital, we developed a data-correction algorithm using the theory of projections. Our method converts clinical knowledge about dependencies in vitals and clinical labs into (non-convex) mixed-integer constraints to model the space of feasible data P, and correct EMR data by computing projections on these non-convex sets. Our results show improved sepsis prediction accuracy, 6 hours before onset, improving upon the current state-of-the-art methods. We are currently exploring incorporating LLMs into such predictive pipelines, and speeding up the projections.

We are further exploring bias in organ transplantation pipelines and potential interventions, in collaboration with MGH.

  • Improving Clinical Decision Support through Interpretable Machine Learning and Error Correction in Electronic Health Records – Mehak Arora, Hassan Mortagy, Nate Dwarshius, Jeffrey Wang, Philip Yang, Swati Gupta, Andre Holder, Rishikesan Kamaleswaran, Journal of the American Medical Informatics Association (JAMIA) 2025.

(d) Mixed-Fidelity Data with Cost: AI is rapidly gainly the monopoly in making decisions across all applications that humans interact with today. AI is not only able to process large amounts of data, but is also able to generalize patterns and integrate into more complex decision-making pipelines. However, AI also has the power to generate fake information, combine data sources with low-quality and high-quality signals, and mimic human interactions with decision systems. This raises a massive challenge in terms of navigating fake data, discounting fake sources of information when training models, and learning from behaviors that look “human” but are machine-generated. A big challenge is to mitigate the propagation of faulty decisions and fake information to the extent possible, without violating the rights of citizens. This not only opens a huge opportunity and market for high-quality data but also calls for intelligent systems that can detect fake and low-quality signals. It is critical that this challenge be addressed, so we have reliability in automated decisions and are not vulnerable to fake data attacks. The AI models we build are as good as the data sources that are used. In this question, we are exploring various questions on understanding the trade-offs between cost, signal and regret in online learning. We are currently exploring these questions in the context of mental health and human-AI augmented systems.

Theme 2. Multi-Criteria Decision-Making with Competing Objectives: This theme considers problems where there are multiple objectives we care about – these might come from various stakeholder preferences, notions of fairness, statistical metrics in machine learning settings, regret guarantees, computational efficiency, variance in decisions, etc.

(a) Portfolios: The theory of multi-criteria optimization to simultaneously optimize for efficiency and fairness is not new. However, it is often unclear how to justify selection of a specific notion of fairness to achieve balance or equity in resource allocation. In the absence of equity-inducing metrics, a resource allocation problem driven by efficiency can over-allocate to certain demographics, while neglecting other groups. To explore this idea, we introduced the notion of solution portfolios: given a class C of fair objectives, find a small set of solutions S so that for any given function f(.) in C, there exists a solution s in S which is a good approximation for optimal under function f(.).

For various combinatorial problems, we are interested in mathematically quantifying the trade-off between the size of the portfolio, and approximation factors possible for various classes of fair objectives C. Constructing these portfolios gives important actionable insight to decision-makers, as they can bypass the choice of the model, but instead focus on the properties of the solutions in the portfolio. Our work on facility location attempts to answer the policy question on developing infrastructure to reduce the number of medical deserts in the US (e.g., explore our app here). We extended this concept to the space of reinforcement learning policies, which allows us to navigate stakeholder preferences in episodic RL, human-AI collaboration through transparent dashboards, and interface with LLMs in policy design.

  • Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning, Cheol Woo Kim, Jai Moondra, Shresth Verma, Madeleine Pollack, Lingkai Kong, Milind Tambe, Swati Gupta, under submission, 2025. arxiv
  • Balancing Notions of Equity: Approximation Algorithms for Fair Portfolio of Solutions in Combinatorial Optimization – Swati Gupta, Jai Moondra, Mohit Singhh, SODA 2025, arxiv.
  • Which Lp norm is the fairest? Approximations for fair facility location across all “p” – Swati Gupta, Jai Moondra, Mohit Singh. EC 2023, arxiv.
  • Too many fairness metrics: Is there a solution? Equity across Demographic Groups for the Facility Location Problem – Swati Gupta, Akhil Jalan, Gireeja Ranade, Helen Yang, Simon Zhuang, Fields Institute Communication Series (forthcoming) and EDSC 2020

(b) Trajectory-Constrained Online Learning: We are interested in developing the theory of online decision-making and iterative optimization, by incorporating trajectory-constraints on the iterates of the algorithm. This helps bring in the perspective of fairness of decisions that are taken over a period of time. These decisions often interact with people, who feel mistreated when they receive decisions that change over time. We considered how a popular notion of individual fairness – that similar individuals should be treated similarly – and proposed an extension over time that generalizes concepts like stare-decisis (in law), markdowns (in retail), and whataboutisms (in politics).

In the context of demand learning for multiple segments of customers, with inter-segment constraints such as price offered to students should be at most the price offered to the general population, we showed that there is no gap in achievable regret rates with and without monotonicity constraints, for smooth and strongly convex functions, while respecting a certain class of inter-segment constraints. These questions are important also in the context of platforms like Upwork, where there is a need to balance opportunities while ensuring that the platform thrives. We explored new algorithmic techniques for allocating jobs, given different learning and performance over different gig economy workers.

  • Algorithmic Challenges in Ensuring Fairness at the Time of Decision with Salem and Kamble, WINE 2022, and (minor rev) Operations Research 2025.
  • Temporal Fairness in Online Decision-Making, Swati Gupta, Jad Salem, Vijay Kamble, Book Chapter in Ethics in Artificial Intelligence: Bias, Fairness and Beyond, 2023.
  • Individual Fairness in Hindsight, Swati Gupta, Vijay Kamble, Journal of Machine Learning Research 2021.
  • Group-Fair Online Allocation in Continuous Time with Semih Cayci, Swati Gupta, Atilla Eryilmaz, NeurIPS 2020.

(c) Districting and Gerrymandering: To understand fair resource allocation in the context of public safety and representation, we looked into districting problems: finding balanced and continuous partitions of a given city so that the parts of the partition are compact. There exist many graph partitioning methods previously, however, these do not guarantee contiguity. In fact, even existing heuristics for continuous graph partitioning were not able to compute a good districting for balancing 911 call workloads for the South Fulton City due to size of the instances. We explored multi-criteria combinatorial approximation methods to find good districting plans for unweighted grid graphs, and heuristic extensions of these to (stochastic) weighted instances. This allowed our algorithms to scale to large instances, and one of the plans from this work was implemented in the SF City. We also looked into gerrymandering in Georgia’s Congressional and state districting plans, and showed that the current plan of Georgia is largely unresponsive to the changing opinions of the electorate (i.e., there is no point of holding elections in GA).

  • Balanced Districting on Grid Graphs with Provable Compactness and Contiguity –
    Cyrus Hettle, Shixiang Zhu, Swati Gupta, Yao Xie, FORC 2021
  • Mathematically Quantifying Non-responsiveness of the 2021 Georgia Congressional Districting Plan – Zhanzhan Zhao, Cyrus Hettle, Swati Gupta, Jonathan C. Mattingly, Dana Randall, Gregory J. Herschlag, EAAMO 2022

(d) Navigating ML and OR Pipelines: We have been exploring various multi-criteria decision-making scenarios in supply chains and power systems. There are many interesting connections between data privacy, algorithmic fairness and reliability in networks under environmental disruptions such as wildfires near transmission networks, and other local grid failures. The overarching question about the interactions of societal operations research with the census data, where OR systems interact with noisy data and ML pipelines provide a massive opportunity for AI and urgent need for reliability and resilience. We are exploring this from various perspectives of robust, stochastic optimization, multi-criteria optimization and policy making.

  • Equitably allocating wildfire resilience investments for power grids: The curse of aggregation and vulnerability indices – Madeleine Pollack, Ryan Piansky, Swati Gupta and Daniel Molzahn, Applied Energy, 2025
  • Fair and Reliable Reconnections for Temporary Disruptions in Electric Distribution Networks – Swati Gupta, Cyrus Hettle and Daniel Molzahn, INFORMS Journal on Computing 2025
  • Generating clusters for urban logistics in hyperconnected networks with Hettle, Faugere, Kwon, Montreuil International Physical Internet Conference IPIC 2021
  • Robust Look-ahead Three-phase Balancing of Uncertain Distribution Loads – Xinbo Geng, Swati Gupta, Le Xie, HICCS 2019
  • Fairness in Inventory Routing with Michael Wang, NeurIPS Workshop on Ethical, Social and Governance Issues in AI, 2018

B. Efficient Optimization and Learning

For scalable structured optimization, we have pushed the boundaries of running time and solution quality by (i) bridging discrete and continuous optimization, (ii) developing novel approximation methods, often multi-criteria, and (iii) exploiting submodularity and combinatorics of the solution space. This includes faster algorithms for structured convex optimization, new ways of rounding relaxed solutions, reusing information from previous subproblems in iterative methods, and improved the quality of solutions in terms of approximation factors. Further, we are interested in new primitives for computation, and bringing machine learning within optimization subroutines for scaling to next-generation applications.

(a) Bridging Discrete and Continuous Optimization: There are significant gaps in our knowledge when it comes to convex optimization over combinatorial polytopoes, even though these polytopes arise frequently in many applications. This is simply due to the fact that the theory of discrete algorithms has evolved largely independently compared to general continuous optimization methods. We settled Bach’s conjecture by exploiting duality of convex minimization over submodular polyhedra (spotlight in NeurIPS), gave the first pyramidal width independent bound for Frank-Wolfe algorithms by analyzing descent directions for constrained optimization, and reduced the time for line search in submodular polyhedra by analyzing the discrete Newton method. We are rapidly developing the theory of combinatorial polytopes using continuous optimization perspectives.

  • Walking in the Shadow: A New Perspective on Descent Directions for Constrained Minimization with Mortagy, Pokutta, NeurIPS 2020, (major rev) Mathematics of Operations Research.
  • Electrical Flows over Spanning Trees with Khodabhaksh, Mortagy, Nikolova Mathematical Programming B, 2022
  • Reusing Combinatorial Structure: Faster Iterative Projections over Submodular Base Polytopes with Moondra, Mortagy NeurIPS 2021
  • Limited Memory Kelley’s Method Converges for Composite Convex and Submodular Objectives with Udell, Zhou NeurIPS 2018 (spotlight)
  • Newton’s Method for Parametric Submodular Function Minimization with Michel Goemans, Patrick Jaillet, IPCO 2017
  • Combinatorial Structure in Online and Convex Optimization, PhD Thesis, MIT 2017
  • Solving Combinatorial Games using Products, Projections and Lexicographically Optimal Bases with Goemans, Jaillet NeurIPS Workshop on Optimization for ML, 2016

(b) Data-driven Optimization and AI: If we have solved an optimization problem already, then solving this on a perturbed dataset, or a related “close-by” optimization should be easier. Most first-order methods solve as a black-box (e.g., linear optimizations in Frank-Wolfe variants, projections in mirror descent variants) perturbations of the same problem, during the process of the overarching optimization. We have developed techniques to reuse combinatorial information (e.g., tight inequalities for submodular polytopes) from previous solves to exploit in the current iterations, resulting in a 2-3 orders of magnitude improvement in the running time. We are critically exploiting submodularity and combinatorial structure in such problems. This direction further lends well to augmented ML algorithms, with provable structure from previous resolves. Can optimization algorithms be speed up using AI, and can AI be speeded up using optimization subroutines? With growing scale of applications that span various constrained decision sets, there is a massive need for the two areas to work collaboratively. We are interested in exploring connections ranging from learning which algorithm works best when to understanding how AI subroutines can be used to guide optimization solvers faster.

  • TACOS: Topology-Aware Collective Algorithm Synthesizer for Distributed Machine Learning with Won, Elavazhagan, Srinivasan, Krishna, MICRO 2024
  • Reusing Combinatorial Structure: Faster Iterative Projections over Submodular Base Polytopes with Moondra, Mortagy, NeurIPS 2021
  • Computational Comparison of Metaheuristics with Silberholz, Golden, Wang, Handbook of Metaheuristics, Springer, 2019
  • What works best when? A Framework for Systematic Heuristic Evaluation with Silberholz, Dunning, INFORMS Journal on Computing, 2018

(c) Approximation Algorithms: We are interested in pushing the boundaries of provable approximations for various combinatorial problems, often motivated by practice. We have developed new approximations for the network reconfiguration problem (from power systems) using kernel-based alpha-point rounding, minimum linear ordering problems, traveling salesman problem, districting, electrical flows over networks, scheduling, ordered set cover, all-norm TSP, facility location, k-clustering, to name a few.

  • Balancing Notions of Equity: Approximation Algorithms for Fair Portfolio of Solutions in Combinatorial Optimization – with Moondra, Singh, SODA 2025, arxiv.
  • Which Lp norm is the fairest? Approximations for fair facility location across all “p” – with Moondra, Singh. EC 2023, arxiv.
  • Hardness and Approximation of Submodular Minimum Linear Ordering Problems, with Farhadi, Sun, Tetali, Wigal Mathematical Programming 2023
  • Electrical Flows over Spanning Trees with Khodabhaksh, Mortagy, Nikolova, Mathematical Programming B, 2022
  • A 4/3 approximation for TSP on cubic 3-edge-connected graph with Agarwal, Garg, Operations Research Letters, 2018

(d) Quantum Optimization: With the growing computational demands of our era, we are looking into new primitives in optimization. With support from a $9.2M DARPA grant for delving into hybrid classical-quantum methods, we are one of the first research groups in the world to bring the concept of warm-starts in quantum optimization. We are currently exploring other classical techniques that can be used to speed up hybrid quantum algorithms, such as sparsification of problem instances, decomposition, warm-starts for other combinatorial problems, incorporation of constraints, etc.

  • Promise of Graph Sparsification and Decomposition for Noise Reduction in QAOA: Analysis for Trapped-Ion Compilations – with Moondra, Mohler, Lotshaw (revision) Quantum 2025
  • Strategies for running the QAOA at hundreds of qubits – with Augustino, Cain, Farhi, Gutmann, Ranard, Tang and Van Kirk
  • Quantum Optimization: Potential, Challenges, and the Path Forward – with 40+ authors, Nature Reviews Physics 2024
  • Classically-inspired Mixers for QAOA Beat Goemans-Williamson’s Max-Cut at Low Circuit Depths, with Tate, Moondra, Gard, Mohler Quantum 2023
  • Generating Target Graph Couplings for QAOA from Native Quantum Hardware Couplings with Rajakumar, Moondra and Herold, Physical Review A, 2022
  • Bridging Classical and Quantum using SDP initialized warm-starts for QAOA with Tate, Farhadi, Herold and Mohler ACM Transactions of Quantum Computing, 2022