About

Kirthevasan Kandasamy is an assistant professor in the Department of Computer Sciences at the University of Wisconsin-Madison, with an affiliation in the Department of Statistics. His research focuses on the intersection of machine learning and game theory. He is a recipient of an NSF CAREER Award in 2025. Prior to his current position, he was a postdoctoral scholar at the RISE Lab at the University of California, Berkeley, where he worked with Ion Stoica, Mike Jordan, and Joey Gonzalez. He completed his PhD in Machine Learning at Carnegie Mellon University under the co-advisement of Jeff Schneider and Barnabas Poczos. During his PhD, he was supported by a Facebook fellowship (2017/18), a Siebel scholarship (2017/18), and a CMU Presidential fellowship (2015/16). Before attending Carnegie Mellon University, he earned a B.Sc in Electronics & Telecommunications Engineering at the University of Moratuwa, Sri Lanka.

Research topics

• Computer Science
• Artificial Intelligence
• Machine Learning
• Parallel computing
• Operating system
• Distributed computing
• Physics

Selected publications

• Pairwise Exchanges of Freely Replicable Goods with Negative Externalities

  arXiv (Cornell University) · 2026-03-12

  articleOpen accessSenior author

  We study a setting where a set of agents engage in pairwise exchanges of freely replicable goods (e.g., digital goods such as data), where two agents grant each other a copy of a good they possess in exchange for a good they lack. Such exchanges introduce a fundamental tension: while agents benefit from acquiring additional goods, they incur negative externalities when others do the same. This dynamic typically arises in real-world scenarios where competing entities may benefit from selective collaboration. For example, in a data sharing consortium, pharmaceutical companies might share (copies of) drug discovery data, when the value of accessing a competitor's data outweighs the risk of revealing their own. In our model, an altruistic central planner wishes to design an exchange protocol (without money), to structure such exchanges between agents. The protocol operates over multiple rounds, proposing sets of pairwise exchanges in each round, which agents may accept or reject. We formulate three key desiderata for such a protocol: (i) individual rationality: agents should not be worse off by participating in the protocol; (ii) incentive-compatibility: agents should be incentivized to share as much as possible by accepting all exchange proposals by the planner; (iii) stability: there should be no further mutually beneficial exchanges upon termination. We design an exchange protocol for the planner that satisfies all three desiderata. While the above desiderata are inspired by classical models for exchange, free-replicability and negative externalities necessitate novel and nontrivial reformalizations of these goals. We also argue that achieving Pareto-efficient agent utilities -- often a central goal in exchange models without externalities -- may be ill-suited in this setting.

  PublisherOA PDF

• Pairwise Exchanges of Freely Replicable Goods with Negative Externalities

  arXiv (Cornell University) · 2026-03-12

  preprintOpen accessSenior author

  We study a setting where a set of agents engage in pairwise exchanges of freely replicable goods (e.g., digital goods such as data), where two agents grant each other a copy of a good they possess in exchange for a good they lack. Such exchanges introduce a fundamental tension: while agents benefit from acquiring additional goods, they incur negative externalities when others do the same. This dynamic typically arises in real-world scenarios where competing entities may benefit from selective collaboration. For example, in a data sharing consortium, pharmaceutical companies might share (copies of) drug discovery data, when the value of accessing a competitor's data outweighs the risk of revealing their own. In our model, an altruistic central planner wishes to design an exchange protocol (without money), to structure such exchanges between agents. The protocol operates over multiple rounds, proposing sets of pairwise exchanges in each round, which agents may accept or reject. We formulate three key desiderata for such a protocol: (i) individual rationality: agents should not be worse off by participating in the protocol; (ii) incentive-compatibility: agents should be incentivized to share as much as possible by accepting all exchange proposals by the planner; (iii) stability: there should be no further mutually beneficial exchanges upon termination. We design an exchange protocol for the planner that satisfies all three desiderata. While the above desiderata are inspired by classical models for exchange, free-replicability and negative externalities necessitate novel and nontrivial reformalizations of these goals. We also argue that achieving Pareto-efficient agent utilities -- often a central goal in exchange models without externalities -- may be ill-suited in this setting.

  PublisherDOI

• Learning from Biased and Costly Data Sources: Minimax-optimal Data Collection under a Budget

  Open MIND · 2026-02-19

  preprintSenior author

  Data collection is a critical component of modern statistical and machine learning pipelines, particularly when data must be gathered from multiple heterogeneous sources to study a target population of interest. In many use cases, such as medical studies or political polling, different sources incur different sampling costs. Observations often have associated group identities (for example, health markers, demographics, or political affiliations) and the relative composition of these groups may differ substantially, both among the source populations and between sources and target population. In this work, we study multi-source data collection under a fixed budget, focusing on the estimation of population means and group-conditional means. We show that naive data collection strategies (e.g. attempting to "match" the target distribution) or relying on standard estimators (e.g. sample mean) can be highly suboptimal. Instead, we develop a sampling plan which maximizes the effective sample size: the total sample size divided by $D_{χ^2}(q\mid\mid\overline{p}) + 1$, where $q$ is the target distribution, $\overline{p}$ is the aggregated source distribution, and $D_{χ^2}$ is the $χ^2$-divergence. We pair this sampling plan with a classical post-stratification estimator and upper bound its risk. We provide matching lower bounds, establishing that our approach achieves the budgeted minimax optimal risk. Our techniques also extend to prediction problems when minimizing the excess risk, providing a principled approach to multi-source learning with costly and heterogeneous data sources.

  DOI

• Constrained Best Arm Identification with Tests for Feasibility

  Proceedings of the AAAI Conference on Artificial Intelligence · 2026-03-14

  articleOpen accessSenior author

  Best arm identification (BAI) aims to identify the highest- performance arm among a set of K arms by collecting stochastic samples from each arm. In real-world problems, the best arm needs to satisfy additional feasibility constraints. While there is limited prior work on BAI with feasibility constraints, they typically assume the performance and con- straints are observed simultaneously on each pull of an arm. However, this assumption does not reflect most practical use cases, e.g., in drug discovery, we wish to find the most potent drug whose toxicity and solubility are below certain safety thresholds. These safety experiments can be conducted separately from the potency measurement. Thus, this requires de- signing BAI algorithms that not only decide which arm to pull but also decide whether to test for the arm’s performance or feasibility. In this work, we study feasible BAI which allows a decision-maker to choose a tuple (i, ℓ), where i ∈ [K] de- notes an arm and ℓ denotes whether she wishes to test for its performance (ℓ = 0) or any of its N feasibility constraints (ℓ ∈ [N ]). We focus on the fixed confidence setting, which is to identify the feasible arm with the highest performance, with a probability of at least 1 − δ. We propose an efficient algorithm and upper-bound its sample complexity, showing our algorithm can naturally adapt to the problem’s difficulty and eliminate arms by worse performance or infeasibility, whichever is easier. We complement this upper bound with a lower bound showing that our algorithm is asymptotically (δ → 0) optimal. Finally, we empirically show that our algorithm outperforms other state-of-the-art BAI algorithms in both synthetic and real-world datasets.

  PublisherDOI

• Learning from Biased and Costly Data Sources: Minimax-optimal Data Collection under a Budget

  ArXiv.org · 2026-02-19

  articleOpen accessSenior author

  Data collection is a critical component of modern statistical and machine learning pipelines, particularly when data must be gathered from multiple heterogeneous sources to study a target population of interest. In many use cases, such as medical studies or political polling, different sources incur different sampling costs. Observations often have associated group identities (for example, health markers, demographics, or political affiliations) and the relative composition of these groups may differ substantially, both among the source populations and between sources and target population. In this work, we study multi-source data collection under a fixed budget, focusing on the estimation of population means and group-conditional means. We show that naive data collection strategies (e.g. attempting to "match" the target distribution) or relying on standard estimators (e.g. sample mean) can be highly suboptimal. Instead, we develop a sampling plan which maximizes the effective sample size: the total sample size divided by $D_{χ^2}(q\mid\mid\overline{p}) + 1$, where $q$ is the target distribution, $\overline{p}$ is the aggregated source distribution, and $D_{χ^2}$ is the $χ^2$-divergence. We pair this sampling plan with a classical post-stratification estimator and upper bound its risk. We provide matching lower bounds, establishing that our approach achieves the budgeted minimax optimal risk. Our techniques also extend to prediction problems when minimizing the excess risk, providing a principled approach to multi-source learning with costly and heterogeneous data sources.

  PublisherOA PDF

• Agora: Bridging the GPU Cloud Resource-Price Disconnect

  ArXiv.org · 2025-09-26

  preprintOpen access

  The historic trend of Moore's Law, which predicted exponential growth in computational performance per dollar, has diverged for modern Graphics Processing Units (GPUs). While Floating Point Operations per Second (FLOPs) capabilities have continued to scale economically, memory bandwidth has not, creating a significant price-performance disconnect. This paper argues that the prevailing time-based pricing models for cloud GPUs are economically inefficient for bandwidth-bound workloads. These models fail to account for the rising marginal cost of memory bandwidth, leading to market distortions and suboptimal hardware allocation. To address this, we propose a novel feature-based pricing framework that directly links cost to resource consumption, including but not limited to memory bandwidth. We provide a robust economic and algorithmic definition of this framework and introduce Agora, a practical and secure system architecture for its implementation. Our implementation of Agora shows that a 50us sampling provides nearly perfect pricing as what ideal sampling would provide - losing only 5\% of revenue. 10us sampling is even better result in 2.4\% loss. Modern telemetry systems can already provide this rate of measurement, and our prototype implementation shows the system design for feature-based pricing is buildable. Our evaluation across diverse GPU applications and hardware generations empirically validates the effectiveness of our approach in creating a more transparent and efficient market for cloud GPU resources.

  PublisherOA PDFDOI

• Constrained Best Arm Identification with Tests for Feasibility

  ArXiv.org · 2025-11-12

  preprintOpen accessSenior author

  Best arm identification (BAI) aims to identify the highest-performance arm among a set of $K$ arms by collecting stochastic samples from each arm. In real-world problems, the best arm needs to satisfy additional feasibility constraints. While there is limited prior work on BAI with feasibility constraints, they typically assume the performance and constraints are observed simultaneously on each pull of an arm. However, this assumption does not reflect most practical use cases, e.g., in drug discovery, we wish to find the most potent drug whose toxicity and solubility are below certain safety thresholds. These safety experiments can be conducted separately from the potency measurement. Thus, this requires designing BAI algorithms that not only decide which arm to pull but also decide whether to test for the arm's performance or feasibility. In this work, we study feasible BAI which allows a decision-maker to choose a tuple $(i,\ell)$, where $i\in [K]$ denotes an arm and $\ell$ denotes whether she wishes to test for its performance ($\ell=0$) or any of its $N$ feasibility constraints ($\ell\in[N]$). We focus on the fixed confidence setting, which is to identify the feasible arm with the highest performance, with a probability of at least $1-δ$. We propose an efficient algorithm and upper-bound its sample complexity, showing our algorithm can naturally adapt to the problem's difficulty and eliminate arms by worse performance or infeasibility, whichever is easier. We complement this upper bound with a lower bound showing that our algorithm is \textit{asymptotically ($δ\rightarrow 0$) optimal}. Finally, we empirically show that our algorithm outperforms other state-of-the-art BAI algorithms in both synthetic and real-world datasets.

  PublisherOA PDFDOI

• Balancing Performance and Costs in Best Arm Identification

  ArXiv.org · 2025-05-26

  preprintOpen accessSenior author

  We consider the problem of identifying the best arm in a multi-armed bandit model. Despite a wealth of literature in the traditional fixed budget and fixed confidence regimes of the best arm identification problem, it still remains a mystery to most practitioners as to how to choose an approach and corresponding budget or confidence parameter. We propose a new formalism to avoid this dilemma altogether by minimizing a risk functional which explicitly balances the performance of the recommended arm and the cost incurred by learning this arm. In this framework, a cost is incurred for each observation during the sampling phase, and upon recommending an arm, a performance penalty is incurred for identifying a suboptimal arm. The learner's goal is to minimize the sum of the penalty and cost. This new regime mirrors the priorities of many practitioners, e.g. maximizing profit in an A/B testing framework, better than classical fixed budget or confidence settings. We derive theoretical lower bounds for the risk of each of two choices for the performance penalty, the probability of misidentification and the simple regret, and propose an algorithm called DBCARE to match these lower bounds up to polylog factors on nearly all problem instances. We then demonstrate the performance of DBCARE on a number of simulated models, comparing to fixed budget and confidence algorithms to show the shortfalls of existing BAI paradigms on this problem.

  PublisherOA PDFDOI

• A Cramér-von Mises Approach to Incentivizing Truthful Data Sharing

  arXiv (Cornell University) · 2025-06-08

  preprintOpen accessSenior author

  Modern data marketplaces and data sharing consortia increasingly rely on incentive mechanisms to encourage agents to contribute data. However, schemes that reward agents based on the quantity of submitted data are vulnerable to manipulation, as agents may submit fabricated or low-quality data to inflate their rewards. Prior work has proposed comparing each agent's data against others' to promote honesty: when others contribute genuine data, the best way to minimize discrepancy is to do the same. Yet prior implementations of this idea rely on very strong assumptions about the data distribution (e.g. Gaussian), limiting their applicability. In this work, we develop reward mechanisms based on a novel, two-sample test inspired by the Cramér-von Mises statistic. Our methods strictly incentivize agents to submit more genuine data, while disincentivizing data fabrication and other types of untruthful reporting. We establish that truthful reporting constitutes a (possibly approximate) Nash equilibrium in both Bayesian and prior-agnostic settings. We theoretically instantiate our method in three canonical data sharing problems and show that it relaxes key assumptions made by prior work. Empirically, we demonstrate that our mechanism incentivizes truthful data sharing via simulations and on real-world language and image data.

  PublisherOA PDFDOI

• Incentivizing Truthful Data Contributions in a Marketplace for Mean Estimation

  ArXiv.org · 2025-02-22

  preprintOpen accessSenior author

  We study a data marketplace where a broker intermediates between buyers, who seek to estimate the mean \(μ\) of an unknown normal distribution \(\Ncal(μ, σ^2)\), and contributors, who can collect data from this distribution at a cost. The broker delegates data collection work to contributors, aggregates reported datasets, sells it to buyers, and redistributes revenue as payments to contributors. We aim to maximize welfare or profit under key constraints: individual rationality for buyers and contributors, incentive compatibility (contributors are incentivized to comply with data collection instructions and truthfully report the collected data), and budget balance (total contributor payments equals total revenue). We first compute welfare/profit-optimal prices under truthful reporting; however, to incentivize data collection and truthful data reporting, we adjust them based on discrepancies in contributors' reported data. This yields a Nash equilibrium (NE) where the two lowest-cost contributors collect all data. We complement this with two hardness results: \emph{(i)} no nontrivial dominant-strategy incentive-compatible mechanism exists in this problem, and \emph{(ii)} no mechanism outperforms ours in a NE.

  PublisherOA PDFDOI

Frequent coauthors

• Barnabás Póczos

  59 shared

• Jeff Schneider

  45 shared

• Joseph E. Gonzalez

  Berkeley College

  23 shared

• Ion Stoica

  23 shared

• Gautam Dasarathy

  23 shared

• Michael I. Jordan

  19 shared

• Junier B. Oliva

  19 shared

• Akshay Krishnamurthy

  13 shared

Labs

• Kandasamy Research LabPI

Education

• B.S., Electronics & Telecommunications Engineering

  University of Moratuwa

• Ph.D.

  Carnegie Mellon University

• M.S.

  University of California, Berkeley

Awards & honors

• NSF CAREER Award