Research & Publications

  • Efficient Uncertainty Quantification of Bagging via the Cheap Bootstrap
    Roy Chowdhury, A., Lam, H.
    • Developed a lightweight bootstrap framework that provides asymptotically exact confidence intervals for bagged predictors using a small, fixed number of resamples as bag size grows.
    • Applicable to Random Forests and bagged estimators in stochastic optimization.
    Preliminary version appeared in Proceedings of the 2025 Winter Simulation Conference (WSC).
    Best Student Paper Award, NESS 2026.
  • Quantification and Decomposition of Uncertainty Using Sliced-Normal Distribution: With Applications to NASA Data
    Roy Chowdhury, A., Crespo, L. G., Lam, H.
    • Developed a scalable framework for modeling complex multivariate distributions with Sliced-Normal models.
    • Reformulated parameter estimation as a convex positive-semidefinite optimization problem, established universal approximation on compact domains, and introduced a block-assembly method for higher-dimensional NASA loss-of-control flight data.
    [arXiv]
  • Superiority of Naïve Optimization via Stochastic Dominance
    Roy Chowdhury, A., Lam, H.
    • Theoretically established that naive methods such as Empirical Risk Minimization (ERM) can outperform Distributionally Robust Optimization (DRO) and regularization in minimizing worst-case regret under joint data and distributional uncertainty, leveraging generalized symmetry arguments.
    • Demonstrated effectiveness under standard distribution-shift regimes.
    [Slides]
  • Estimation of Rare-Event Probabilities Using Extreme Value Theory
    Roy Chowdhury, A., Lam, H., Long, D.
    • Estimates rare-event probabilities by modeling the tail behavior using the Generalized Pareto Distribution (GPD).
    • Investigates GPD-based methods as a model-agnostic alternative to importance sampling for rare-event simulation.
  • Bayesian Nonparametric Generalization of Tree-Based ML Approaches
    Roy Chowdhury, A., Bhattacharya, S.
    • Proposed a Bayesian nonparametric mixture model of decision trees with a random number of components, using a Dirichlet Process prior to flexibly capture complex decision boundaries.
    M.Stat dissertation work
    [Report] [Slides]
  • Accelerography: Feasibility of Gesture Typing Using Accelerometer
    Roy Chowdhury, A.*, Dalal, A.*, Sen, S.*
    • Designed intuitive motion gestures for the English alphabet, enabling text input by physically moving a phone.
    • Developed and implemented an R-based pipeline to classify these gestures using accelerometer data.
    * Equal contribution
    [arXiv]

Invited Talks

  • NASA Langley Research Centre

    Session: Uncertainty Quantification Seminar (Nov 2024)

    • Title 1: Modelling Multivariate Data using Exponential Polynomials
    • Title 2: Stochastic Optimization: Superiority of Naïve Approaches
  • INFORMS Annual Meeting

    Session: Recent Advances in Data-Driven Optimization (Oct 2024)

    • Title: Robustness vs Statistical Efficiency: Superiority of Naïve Optimization
  • The 37th New England Statistics Symposium

    Session: Recent Advances in Data-Driven Decision-Making (May 2024)

    • Title: Regret Optimality of Empirical Risk Minimization

Early Research & Project Reports

  • Causal Inference Using Directed Acyclic Graphical (DAG) Models (Nov–Dec 2019)
    Roy, S., Dalal, A., Roy Chowdhury, A.
    • Explored the need of studying causal inference as a separate topic, and did a basic review of causal inference using directed acyclic graphs.
    • Performed analysis of causal inference on different datasets, to estimate the underlying causal model.
    • Noted some significant implications for Tea auction mechanism in India.
    [Report]
  • Modelling Summit Success Rate of Mt. Rainier (Nov–Dec 2019)
    Roy, S., Dalal, A., Roy Chowdhury, A., Bakshi, A.
    • Modeled team size as a Poisson mixture and analyzed summit success rates conditioned on team size, achieving strong predictive performance.
    [Slides]