Publications

We study off-policy learning from observational data under covariate shift and local differential privacy (LDP) constraints. Modern policy learning faces a fundamental trilemma: incorporating high-dimensional contexts improves performance but requires pooling external data, which introduces distribution shifts and privacy risks. We formalize the difficulty of this problem using the LDP parameter, the transfer exponent, and a relaxed, localized treatment coverage assumption rather than strict global overlap. Under this framework, we establish minimax lower bounds on the policy sub-optimal gap for both non-private and \varepsilon-LDP settings. To achieve these fundamental limits, we develop a policy learning algorithm grounded in a unified pessimism principle that addresses both covariate shift and privacy-induced uncertainty. By estimating the conditional outcome via nonparametric regression and constructing a lower confidence bound that penalizes underexplored treatments and privacy noise, our approach bypasses the estimation of possibly heterogeneous historical logging policies. We prove that maximizing this pessimistic objective yields a minimax-optimal policy in the non-private setting, fully characterizing the theoretical trade-offs between data dimensionality and distribution shift. We conjecture that this pessimistic policy achieves matching optimality under privacy constraints.

High-dimensional sparse clustering is a combinatorial NP-hard problem that arises from the coupling between cluster assignment and variable selection. We demonstrate that semidefinite programming (SDP) relaxation of K-means is robust to variable over-selection by establishing minimax separation bounds. Leveraging this robustness, we propose a block-coordinate ascent framework that alternates between SDP-based clustering and conservative variable selection. To address the tendency of deterministic greedy methods to become trapped in local optima, we formulate the variable selection step as a bandit problem. Crucially, to reliably evaluate feature utility and generate stochastic rewards for Thompson sampling even under imperfect intermediate cluster assignments, we employ a robust maximum mean discrepancy (MMD) permutation test. This approach introduces adaptive memory by aggregating historical variable-selection outcomes into posterior distributions, and selects features via posterior sampling, enabling stochastic exploration that promotes the inclusion of underexplored features and facilitates escape from local maxima. We establish conditions for consistent variable selection and exact cluster recovery, and extend the method to settings with unknown covariance through a scalable estimation procedure. Synthetic experiments and a real-data application in document clustering demonstrate that the proposed memory-driven randomized approach consistently outperforms state-of-the-art sparse clustering methods.

Motivated by renal imaging studies that combine renogram curves with pharmacokinetic and demographic covariates, we propose Hybrid partial least squares (Hybrid PLS) for simultaneous supervised dimension reduction and regression in the presence of cross-modality correlations. The proposed approach embeds multiple functional and scalar predictors into a unified hybrid Hilbert space and rigorously extends the nonlinear iterative PLS (NIPALS) algorithm. This theoretical development is complemented by a sample-level algorithm that incorporates roughness penalties to control smoothness. By exploiting the rank-one structure of the resulting optimization problem, the algorithm admits a computationally efficient closed-form solution that requires solving only linear systems at each iteration. We establish fundamental geometric properties of the proposed framework, including orthogonality of the latent scores and PLS directions. Extensive numerical studies on synthetic data, together with an application to a renal imaging study, validate these theoretical results and demonstrate the method’s ability to recover predictive structure under intermodal multicollinearity, yielding parsimonious low-dimensional representations.

Federated learning, where user data remains on local devices and only weight updates are transmitted, has become the industry standard for training neural networks on sensitive data. Federated analytics builds on this infrastructure, enabling data science insights without transmitting raw data. However, recent studies show that hijacking weight updates or data summaries can allow recovery of the original raw data. To address this privacy risk, we propose a private A/B testing method that transmits noisy data summaries within the federated learning framework. We begin with multinomial data, introducing private permutation tests using privacy mechanisms like the Laplace mechanism, discrete Laplace mechanism, and Google’s RAPPOR mechanism. We extend our approach to continuous data using binning and analyze uniform separation rates under local differential privacy (LDP). Our tests rigorously control Type I error, satisfy LDP constraints, and achieve minimax separation rates, highlighting the inherent privacy-utility trade-offs in private testing.

Data augmentation and loss function adjustments are two common techniques for imbalanced classification. In cases of extreme class imbalance, it is often standard practice to combine these two approaches. However, determining the optimal oversampling ratio and the degree of asymmetry in the loss function typically relies on heuristics. We further consider the scenario where the minority class consists of subgroups and propose a straightforward method for combining data augmentation and loss function adjustments. This method serves as a sample-based approximation of the population-level asymptotically Bayes optimal oracle procedure.

Functional connectivity refers to the correlations between neural signal time series across different brain regions, in contrast to the physical (structural) connections between them. We are interested in understanding how autism causally affects functional connectivity, and whether this effect differs between sexes. From a statistical perspective, this problem corresponds to estimating heterogeneous treatment effects (HTEs). Because autism spectrum disorder (ASD) is diagnosed more frequently in males than in females, the number of female participants in the autism group is substantially smaller. As a result, female samples contribute less to the HTE estimation algorithm, leading to potentially unstable estimates for females. To address this subgroup imbalance, we employ oversampling using a cluster-aware generative model. The augmented data increase the representation of female samples in the HTE estimation process, thereby improving the stability and reliability of HTE estimates for females. Our results reveal that several brain regions exhibit significant HTEs that were not detectable prior to oversampling. Furthermore, we demonstrate that the identified HTEs are not random noise but have clear biological meaning: they are strongly associated with higher-order cognitive control functions, rather than lower-level sensory processing, and are linked to autism-related genes.

Artillery training inherently poses wildfire risks. Predictive modeling of these wildfires faces two key challenges: the scarcity of wildfire cases and the limited granularity of meteorological data during training. We address the first challenge by augmenting the data using a Gaussian mixture generative model and adjusting the loss function of a support vector machine. To tackle the second challenge, we integrate the Republic of Korea Army (ROKA) dataset with the Korea Meteorological Administration database. Our resulting model achieves a 99% improvement in balanced classification metrics compared to previous models.

To enable high-quality neural recordings in freely moving animals, our materials science team developed a liquid metal–based neural probe that can be directly printed onto the brain, along with elastic electronic circuits. To validate the fidelity of the signals recorded by this neural probe, we statistically assessed the long-term stability of signal clustering and examined the correlations between micro-scale neuronal activity and macro-level brain dynamics during mouse behavioral experiments involving visual and motor stimuli.