SEEDSConference 2025
January 8-11, 2025
University of Southern California, Los Angeles (CA)
University of Southern California, Los Angeles (CA)
| Xihong Lin (Harvard University) | |
| Navigate the Crossroad of Statistics, ML/AI and Domain Science | |
| Abstract: The data science ecosystem encompasses data fairness, scalable statistical and ML/AI methods and tools, interpretable data analysis, and trustworthy decision-making. Rapid advancements in AI have revolutionized data utilization and enabled machines to learn from data more effectively. Statistics, as the science of learning from data while accounting for uncertainty, plays a pivotal role in addressing complex real-world problems and facilitating trustworthy decision-making. In this talk, I will discuss the challenges and opportunities as we navigate the crossroad of statistics and AI, including how to build an end-to-end scalable data science ecosystem, leverage AI/ML-prediction to empower statistical analysis of biobank data, transfer inference for genetic association analysis, build the whole genome variant functional annotation database and portal FAVOR and FAVOR-GPT, and incorporate multi-faceted variant functional annotation to boost power of whole genome sequencing end-to-end analysis. This talk aims to ignite proactive and thought-provoking discussions, foster cross-disciplinary collaboration, and cultivate open-minded approaches to advance scientific discovery. |
| Richard Samworth (University of Cambridge) | |
| How should we do linear regression? | |
| Abstract: In the context of linear regression, we construct a data-driven convex loss function with respect to which empirical risk minimisation yields optimal asymptotic variance in the downstream estimation of the regression coefficients. Our semiparametric approach targets the best decreasing approximation of the derivative of the log-density of the noise distribution. At the population level, this fitting process is a nonparametric extension of score matching, corresponding to a log-concave projection of the noise distribution with respect to the Fisher divergence. The procedure is computationally efficient, and we prove that our procedure attains the minimal asymptotic covariance among all convex M-estimators. As an example of a non-log-concave setting, for Cauchy errors, the optimal convex loss function is Huber-like, and our procedure yields an asymptotic efficiency greater than 0.87 relative to the oracle maximum likelihood estimator of the regression coefficients that uses knowledge of this error distribution; in this sense, we obtain robustness without sacrificing much efficiency. |
| Jianqing Fan (Princeton University) | |
| Spectral Ranking Inferences Based on General Multiway Comparisons | |
| Abstract: This paper studies the performance of the spectral method in the estimation and uncertainty quantification of the unobserved preference scores of compared entities in a general and more realistic setup. Specifically, the comparison graph consists of hyper-edges of possible heterogeneous sizes, and the number of comparisons can be as low as one for a given hyper-edge. Such a setting is pervasive in real applications, circumventing the need to specify the graph randomness and the restrictive homogeneous sampling assumption imposed in the commonly used Bradley-Terry-Luce (BTL) or Plackett-Luce (PL) models. Furthermore, in scenarios where the BTL or PL models are appropriate, we unravel the relationship between the spectral estimator and the Maximum Likelihood Estimator (MLE). We discover that a two-step spectral method, where we apply the optimal weighting estimated from the equal weighting vanilla spectral method, can achieve the same asymptotic efficiency as the MLE. Given the asymptotic distributions of the estimated preference scores, we also introduce a comprehensive framework to carry out both one-sample and two-sample ranking inferences, applicable to both fixed and random graph settings. It is noteworthy that this is the first time effective two-sample rank testing methods have been proposed. Finally, we substantiate our findings via comprehensive numerical simulations and subsequently apply our developed methodologies to perform statistical inferences for statistical journals and movie rankings. (Joint work with Zhipeng Lou, Weichen Wang, and Mengxin Yu) |
| Thomas Berrett (University of Warwick) | |
| Efficient estimation with incomplete data via generalised ANOVA decompositions | |
| Abstract: In this talk I will present recent work (https://arxiv.org/abs/2409.05729) on efficient estimation with incomplete data, covering problems arising in semi-supervised learning, data fusion and missing data literatures. Our task is to estimate simple mean functionals given access to a complete dataset that is supplemented by additional incomplete datasets. In particular, we aim to use the incomplete data to reduce the variance of the naive complete-case estimator, and to characterise the minimal asymptotic risk among all estimators. Results of this type exist for monotonic missingness structures, such as those arising in semi-supervised learning and longitudinal studies, but in this work we consider more general settings. We show that the optimal variance can be expressed through the minimal value of a quadratic optimisation problem over a function space, thus establishing a fundamental link between these estimation problems and the theory of generalised ANOVA decompositions. We introduce an estimator that is proved to attain this minimal risk and to be approximately normally distributed, and use this to construct confidence intervals. |
| Timothy Cannings (University of Edinburgh) | |
| Nonparametric classification with missing data | |
| Abstract: We introduce a new nonparametric framework for classification problems in the presence of missing data. The key aspect of our framework is that the regression function decomposes into an anova-type sum of orthogonal functions, of which some (or even many) may be zero. Working under a general missingness setting, which allows features to be missing not at random, our main goal is to derive the minimax rate for the excess risk in this problem. In addition to the decomposition property, the rate depends on parameters that control the tail behaviour of the marginal feature distributions, the smoothness of the regression function and a margin condition. The ambient data dimension does not appear in the minimax rate, which can therefore be faster than in the classical nonparametric setting. We further propose a new method, called the Hard-thresholding Anova Missing data (HAM) classifier, based on a careful combination of a k-nearest neighbour algorithm and a thresholding step. The HAM classifier attains the minimax rate up to polylogarithmic factors and numerical experiments further illustrate its utility. |
| Jean Feng (UC San Francisco) | |
| Bayesian Concept Bottleneck Models with LLM Priors | |
| Abstract: Concept Bottleneck Models (CBMs) have been proposed as a compromise between white-box and black-box models, aiming to achieve interpretability without sacrificing accuracy. The standard training procedure for CBMs is to predefine a candidate set of human-interpretable concepts, extract their values from the training data, and identify a sparse subset as inputs to a transparent prediction model. However, such approaches are often hampered by the tradeoff between enumerating a sufficiently large set of concepts to include those that are truly relevant versus controlling the cost of obtaining concept extractions. This work investigates a novel approach that sidesteps these challenges: BC-LLM iteratively searches over a potentially infinite set of concepts within a Bayesian framework, in which Large Language Models (LLMs) serve as both a concept extraction mechanism and prior. We show that BC-LLM provides rigorous statistical inference despite imperfections in LLMs and consistently outperforms comparator methods in experiments. |
| Yang Feng (New York University) | |
| Regularized Fine-tuning in Representation Multi-task Learning: Adaptivity and Robustness | |
| Abstract: Representation multi-task learning (MTL) has achieved tremendous success in practice. However, the theoretical understanding of these methods is still lacking. Most existing theoretical works focus on cases where all tasks share the same representation, and claim that MTL almost always improves performance. Nevertheless, as the number of tasks grows, assuming all tasks share the same representation is unrealistic. Furthermore, empirical findings often indicate that a shared representation does not necessarily improve single-task learning performance. In this paper, we aim to understand how to learn from tasks with similar but not exactly the same linear representations, while dealing with outlier tasks. Assuming a known intrinsic dimension, we proposed a penalized empirical risk minimization method and a spectral method that are adaptive to the similarity structure and robust to outlier tasks. Both algorithms outperform single-task learning when representations across tasks are sufficiently similar and the proportion of outlier tasks is small. Moreover, they always perform at least as well as single-task learning, even when the representations are dissimilar. We provided information-theoretic lower bounds to demonstrate that both methods are nearly minimax optimal in a large regime, with the spectral method being optimal in the absence of outlier tasks. Additionally, we introduce a thresholding algorithm to adapt to an unknown intrinsic dimension. We conducted extensive numerical experiments to validate our theoretical findings. |
| Lan Gao (University of Tennessee, Knoxville) | |
| ARK: Robust Knockoffs Inference with Coupling | |
| Abstract: We investigate the robustness of the model-X knockoffs framework with respect to the misspecified or estimated feature distribution. We achieve such a goal by theoretically studying the feature selection performance of a practically implemented knockoffs algorithm, which we name as the approximate knockoffs (ARK) procedure, under the measures of the false discovery rate (FDR) and family wise error rate (FWER). The approximate knockoffs procedure differs from the model-X knockoffs procedure only in that the former uses the misspecified or estimated feature distribution. A key technique in our theoretical analyses is to couple the approximate knockoffs procedure with the model-X knockoffs procedure so that random variables in these two procedures can be close in realizations. We prove that if such coupled model-X knockoffs procedure exists, the approximate knockoffs procedure can achieve the asymptotic FDR or FWER control at the target level. We showcase three specific constructions of such coupled model-X knockoff variables, verifying their existence and justifying the robustness of the model-X knockoffs framework. |
| Jose Angel Sanchez Gomez (UC Riverside) | |
| Detecting hub variables in large Gaussian graphical models | |
| Abstract: In modern scientific applications, identifying small sets of variables in a dataset with a strong influence over the rest is often vital. For example, when studying the gene-expression levels of cancer patients, estimating the most influential genes can be a first step towards understanding underlying gene dynamics and proposing new treatments. A popular approach for representing variable influence is through a Gaussian graphical model (GGM), where each variable corresponds to a node, and a link between two nodes represents relationships among pairs of variables. In a GGM, influential variables correspond to nodes with a high degree of connectivity, also known as hub variables. In this talk, I share a new method for estimating hub variables in GGMs. To this end, we establish a connection between the presence of hubs in a GGM and the concentration of principal component vectors on the hub variables. We provide probabilistic guarantees of convergence for our method, even in high-dimensional data where the number of variables can be arbitrarily large. I will also discuss an application of this new method to a prostate cancer gene-expression dataset, through which we detect several hub genes with close connections to tumor development. |
| Yuqi Gu (Columbia University) | |
| Minimax-Optimal Covariance Projected Spectral Clustering for High-Dimensional Anisotropic Mixtures | |
| Abstract: In mixture models, nonspherical or anisotropic noise within each cluster is widely present in real-world data. We consider both the minimax rate and estimation procedure for clustering under high-dimensional anisotropic mixture models. In high-dimensional settings, we first establish the information-theoretic limits for clustering under Gaussian mixtures. The minimax lower bound unveils an intriguing informational dimension-reduction phenomenon. Motivated by the lower bound, we propose a novel computationally efficient clustering method, the Covariance Projected Spectral Clustering. This method is designed to discern variations in the covariance matrices across different clusters in the lens of a low-dimensional subspace. Its key step is to project the high-dimensional data onto the space spanned by the cluster centers and then use the covariance adjustment in this space to enhance clustering. Theoretically, we establish tight algorithmic upper bounds for our method, both for Gaussian noise with arbitrary dependence and general noise with local dependence. Our theory indicates the minimax-optimality of our method in the Gaussian case and highlights its adaptivity to a broad spectrum of dependent noise. Extensive simulation studies under various noise structures and a high-dimensional genetics dataset analysis demonstrate our method's superior performance. |
| Ana Maria Kenney (UC Irvine) | |
| Distilling Causal Effects: Stable subgroup estimation via distillation trees in causal inference | |
| Abstract: Recent methodological developments have introduced new black-box approaches to better estimate heterogeneous treatment effects; however, these methods fall short of providing interpretable characterizations of the underlying individuals who may be most at risk or benefit most from receiving the treatment, thereby limiting their practical utility. In this work, we introduce a novel method, causal distillation trees (CDT), to estimate interpretable subgroups. CDT allows researchers to fit any machine learning model of their choice to estimate the individual-level treatment effect, and then leverages a simple, second-stage tree-based model to “distill” the estimated treatment effect into meaningful subgroups. As a result, CDT inherits the theoretical guarantees from black-box machine learning models while preserving the interpretability of a simple decision tree. We theoretically characterize the stability of CDT in estimating substantively meaningful subgroups and provide stability-driven diagnostics for researchers to evaluate the quality of the estimated subgroups. We illustrate our proposed method on a randomized controlled trial of antiretroviral treatment for HIV from the AIDS Clinical Trials Group Study 175 and show that CDT out-performs state-of-the-art approaches in constructing stable, clinically relevant subgroups. |
| Jingyi Jessica Li (UC Los Angeles) | |
| SyNPar: A Data-Preservation Framework for High-Power False Discovery Rate Control in High-Dimensional Variable Selection | |
| Abstract: Balancing false discovery rate (FDR) control and statistical power is a fundamental challenge in high-dimensional variable selection. Existing FDR control methods often perturb the original data, either by concatenating knockoff variables or splitting the data, which can compromise power. In this paper, we introduce SyNPar, a novel framework that controls the FDR in high-dimensional variable selection while preserving the integrity of the original data. SyNPar generates synthetic null data using an inference model under the null hypothesis and identifies false positives through a numerical analog of the likelihood ratio test. We provide rigorous theoretical guarantees for FDR control at any desired level and show that SyNPar achieves asymptotically optimal power. The framework is versatile, straightforward to implement, and applicable to a wide range of statistical models, including high-dimensional linear regression, generalized linear models (GLMs), Cox models, and Gaussian graphical models. Through extensive simulations and real-world data applications, we demonstrate that SyNPar consistently outperforms state-of-the-art methods, such as knockoff and data-splitting techniques, in terms of FDR control, statistical power, and computational efficiency. |
| Anna Neufeld (Williams College) | |
| Data thinning to avoid double dipping | |
| Abstract: We refer to the practice of using the same data to fit and validate a model as double dipping. Problems arise when standard statistical procedures are applied in settings that involve double dipping. To circumvent the challenges associated with double dipping, one approach is to fit a model on one dataset, and then validate the model on another independent dataset. When we only have access to one dataset, we typically accomplish this via sample splitting. Unfortunately, in many unsupervised problems, sample splitting does not allow us to avoid double dipping. In this talk, we are motivated by unsupervised problems that arise in the analysis of single cell RNA sequencing data. We first propose Poisson count splitting, which splits a single observation drawn from a Poisson distribution into two independent components. We show that Poisson count splitting allows us to avoid double dipping in unsupervised settings. We next generalize the count splitting framework to a variety of distributions, and refer to the generalized framework as data thinning. Data thinning is a very general alternative to sample splitting that is useful far beyond the context of single-cell RNA sequencing data, and, unlike sample splitting, can be applied in both supervised and unsupervised settings. |
| Snigdha Panigrahi (University of Michigan) | |
| Cross-validation with antithetic Gaussian randomization | |
| Abstract: In this talk, I will introduce a new method for performing cross-validation using “antithetic” Gaussian randomization variables. The randomization variables in our method are drawn from an equicorrelated, degenerate normal distribution. Each pair of these randomization variables is maximally negatively correlated, which is why we describe this randomization scheme as “antithetic”. Inspired by recent data-splitting techniques such as data-fission and data-thinning, the new cross-validation method is well-suited for problems where traditional sample splitting is infeasible. Even in scenarios where traditional sample splitting is possible, our cross-validation method offers a computationally efficient alternative for estimating prediction error. By reducing the amount of randomization in the train data, our method achieves bias levels as small as the standard leave-one-out cross-validation, while requiring only a small number of train-test repetitions---potentially as few as two. A key advantage of our cross-validated estimator is its stable variance, which does not increase even as the bias from estimating the prediction function on the training data approaches zero. In both theory and simulations, we show that this desirable bias-variance property of our cross-validated estimator extends to a wide range of loss functions, including those commonly used in generalized linear models. This is based on joint work with Sifan Liu and Jake Soloff. |
| Nicole Pashley (Rutgers University) | |
| Instrumental Variable Methods for Factorial Experiments with Complex Treatment Uptake | |
| Abstract: There is a well-established literature dealing with noncompliance in treatment-control designs within the potential-outcome framework for causal inference. However, generalizing to experiments with more than two treatment arms and noncompliance remains a challenge with limited exploration. The focus of this talk will be noncompliance in two-level factorial designs. The talk will discuss why this setting is so challenging, propose different assumptions to learn about relevant estimands, and explore identification and inference results under these assumptions. |
| Annie Qu (UC Irvine) | |
| Stage-Aware Learning for Dynamic Treatments | |
| Abstract: Recent advances in dynamic treatment regimes (DTRs) provide powerful optimal treatment searching algorithms, which are tailored to individuals’ specific needs and able to maximize their expected clinical benefits. However, existing algorithms could suffer from insufficient sample size under optimal treatments, especially for chronic diseases involving long stages of decision-making. To address these challenges, we propose a novel individualized learning method which estimates the DTR with a focus on prioritizing alignment between the observed treatment trajectory and the one obtained by the optimal regime across decision stages. By relaxing the restriction that the observed trajectory must be fully aligned with the optimal treatments, our approach substantially improves the sample efficiency and stability of inverse probability weighted based methods. In particular, the proposed learning scheme builds a more general framework which includes the popular outcome weighted learning framework as a special case of ours. Moreover, we introduce the notion of stage importance scores along with an attention mechanism to explicitly account for heterogeneity among decision stages. We establish the theoretical properties of the proposed approach, including the Fisher consistency and finite-sample performance bound. Empirically, we evaluate the proposed method in extensive simulated environments and a real case study for COVID-19 pandemic. |
| Zhimei Ren (University of Pennsylvania) | |
| Confidence on the Focal: Conformal Prediction with Selection-Conditional Coverage | |
| Abstract: Conformal prediction builds marginally valid prediction intervals that cover the unknown outcome of a randomly drawn new test point with a prescribed probability. However, a common scenario in practice is that, after seeing the data, practitioners decide which test unit(s) to focus on in a data-driven manner and seek for uncertainty quantification of the focal unit(s). In such cases, marginally valid conformal prediction intervals may not provide valid coverage for the focal unit(s) due to selection bias. In this talk, I will present a general framework for constructing a prediction set with finite-sample exact coverage conditional on the unit being selected by a given procedure. The general form of our method works for arbitrary selection rules that are invariant to the permutation of the calibration units, and generalizes Mondrian Conformal Prediction to multiple test units and non-equivariant classifiers. We then work out the computationally efficient implementation of our framework for a number of realistic selection rules, including top-K selection, optimization-based selection, selection based on conformal p-values, and selection based on properties of preliminary conformal prediction sets. The performance of our methods is demonstrated via applications in drug discovery and health risk prediction. |
| Alessandro Rinaldo (University of Texas at Austin) | |
| Statistical Inference for Temporal Difference Learning with Linear Function Approximation | |
| Abstract: Policy evaluation is a fundamental task in Reinforcement Learning (RL), with applications in numerous fields, such as clinical trials, mobile health, robotics, and autonomous driving. Temporal Difference (TD) learning and its variants are arguably the most widely used algorithms for policy evaluation with linear approximation. Despite the popularity and practical importance of TD estimators of the parameters of the best linear approximation to the value function, theories and methods for formal statistical inference with finite sample validity in high dimensions remain limited. Consequently, RL practitioners often lack essential statistical tools to guide their decision-making. To address this gap, we develop efficient inference procedures for TD learning-based estimators under linear function approximation in on-policy settings. We obtain improved consistency rates and derive novel high-dimensional Berry-Esseen bounds for the TD estimator under independent samples and Markovian trajectories. Additionally, we propose an online algorithm to construct non-asymptotic confidence intervals for the target parameters. Joint work with Weichen Wu (Voleon) and Yuting Wei (UPenn). |
| Yaniv Romano (Technion) | |
| Robust Conformal Prediction Using Privileged Information | |
| Abstract: This talk introduces a novel method for generating uncertainty sets for ML predictions with guaranteed coverage for the unknown test label in the face of distribution shifts induced by corrupted training data, such as missing or noisy variables. Our approach builds on conformal prediction—a powerful framework for constructing uncertainty sets—but addresses its limitations in scenarios involving such distribution shifts. Central to our method is the use of privileged information (PI)—additional features available only during training that capture valuable insights about the training data, such as the expertise level of annotators, time spent on labeling, or disagreements among annotators. While such PI features are unavailable at test time, they play a crucial role in explaining the distribution shift, paving the way to construct valid and robust uncertainty estimates. I will demonstrate the practical utility of this method through real-world applications, including scenarios with missing labels and, if time permits, unobserved counterfactuals at test time in causal inference tasks. |
| Cynthia Rush (Columbia University) | |
| Is It Easier to Count Communities Than Find Them? | |
| Abstract: Random graph models with community structure have been studied extensively in the literature. For both the problems of detecting and recovering community structure, an interesting landscape of statistical and computational phase transitions has emerged. A natural unanswered question is: might it be possible to infer properties of the community structure (for instance, the number and sizes of communities) even in situations where actually finding those communities is believed to be computationally hard? We show the answer is no. In particular, we consider certain hypothesis testing problems between models with different community structures, and we show (in the low-degree polynomial framework) that testing between two options is as hard as finding the communities. In addition, our methods give the first computational lower bounds for testing between two different “planted” distributions, whereas previous results have considered testing between a planted distribution and an i.i.d. “null” distribution. This is joint work with Fiona Skerman, Alexander S. Wein, and Dana Yang. |
| Damla Senturk (UC Los Angeles) | |
| Modeling intra-individual inter-trial EEG response variability in autism | |
| Abstract: Autism spectrum disorder (autism) is a prevalent neurodevelopmental condition characterized by early emerging impairments in social behavior and communication. EEG represents a powerful and non-invasive tool for examining functional brain differences in autism. Recent EEG evidence suggests that greater intra-individual trial-to-trial variability across EEG responses in stimulus-related tasks may characterize brain differences in autism. Traditional analysis of EEG data largely focuses on mean trends of the trial-averaged data, where trial-level analysis is rarely performed due to low neural signal to noise ratio. We propose to use nonlinear (shape-invariant) mixed effects (NLME) models to study intra-individual inter-trial EEG response variability using trial-level EEG data. By providing more precise metrics of response variability, this approach could enrich our understanding of neural disparities in autism and potentially aid the identification of objective markers. The proposed multilevel NLME models quantify variability in the signal’s interpretable and widely recognized features (e.g., latency and amplitude) while also regularizing estimation based on noisy trial-level data. Even though NLME models have been studied for more than three decades, existing methods cannot scale up to large data sets. We propose computationally feasible estimation and inference methods via the use of a novel minorization-maximization (MM) algorithm. Extensive simulations are conducted to show the efficacy of the proposed procedures. Applications to data from a large national consortium find that children with autism have larger intra-individual inter-trial variability in P1 latency in a visual evoked potential (VEP) task, compared to their neurotypical peers. |
| Jane-Ling Wang (UC Davis) | |
| Hypothesis testing for black-box survival model | |
| Abstract: Deep learning has become enormously popular in the analysis of complex data, including event time measurements with censoring. To date, deep survival methods have mainly focused on prediction. Such methods are scarcely used in matters of statistical inference such as hypothesis testing. Due to their black-box nature, deep-learned outcomes lack interpretability which limits their use for decision-making in biomedical applications. Moreover, conventional tests fail to produce reliable type I errors due to the ability of deep neural networks to learn the data structure under the null hypothesis even if they search over the full space. This talk provides testing methods for survival models and demonstrates its use in the nonparametric Cox model, where the nonparametric link function is modeled via a deep neural network. To perform hypothesis testing, we utilize sample splitting and cross-fitting procedures to get neural network estimators and construct the test statistic. These procedures enable us to propose a new significance test to examine the association of certain covariates with event times. We show that our test statistic converges to a normal distribution under the null hypothesis and establish its consistency, in terms of the Type II error, under the alternative hypothesis. Numerical simulations and a real data application demonstrate the usefulness of the proposed test. |
| Lan Wang (University of Miami) | |
| Differentially Private Quantile Regression with Applications to Inventory Policy Learning | |
| Abstract: We introduce a novel approach for privacy-preserving quantile regression within the f -differential privacy framework, an extension of the classical (ϵ, δ)-differential privacy with several appealing properties. A key challenge is the nonsmoothness of the quantile loss function, which sets it apart from existing work on privacy-preserving algorithms in other settings. We develop a clipped noisy gradient descent algorithm based on convolution smoothing for quantile regression, and show that the algorithm provides privacy guarantees and desirable statistical precision. We derive finite-sample high-probability bounds for parameter estimation and regret analysis. Our bound aligns with that for strongly convex and smooth loss function. We apply the approach to the data-driven newsvendor problem with features and show that we attain a faster excess population risk bound compared to that obtained from an indiscriminate application of existing results for general nonsmooth convex loss. Our numerical experiments demonstrate that the proposed new method can achieve desirable privacy protection with a marginal increase in cost. (Joint work with Tuoyi Zhao and Wenxin Zhou). |
| Miaoyan Wang (University of Wisconsin-Madison) | |
| Application and Methods for Structured Tensor Learning | |
| Abstract: High-order tensor datasets pose common challenges in applications such as recommendation systems, neuroimaging, and social networks. In this work, we introduce two approaches for learning with structured tensors: tensor block models for higher-order clustering and sign-series models for tensor denoising. These approaches provide a lens into the unique properties of tensor analysis. We establish statistical and computational efficiency results for each method. Additionally, we present polynomial-time algorithms with guaranteed efficiency. The effectiveness of our methods is demonstrated through applications to neuroimaging data analysis and social network analysis. |
| Yi Yu (University of Warwick) | |
| Optimal estimation in private distributed functional data analysis | |
| Abstract: We systematically investigate the preservation of differential privacy in functional data analysis, beginning with functional mean estimation and extending to varying coefficient model estimation. Our work introduces a distributed learning framework involving multiple servers, each responsible for collecting several sparsely observed functions. This hierarchical setup introduces a mixed notion of privacy. Within each function, user-level differential privacy is applied to m discrete observations. At the server level, central differential privacy is deployed to account for the centralised nature of data collection. Across servers, only private information is exchanged, adhering to federated differential privacy constraints. To address this complex hierarchy, we employ minimax theory to reveal several fundamental phenomena: from sparse to dense functional data analysis, from user-level to central and federated differential privacy costs, and the intricate interplay between different regimes of functional data analysis and privacy preservation. To the best of our knowledge, this is the first study to rigorously examine functional data estimation under multiple privacy constraints. Our theoretical findings are complemented by efficient private algorithms and extensive numerical evidence, providing a comprehensive exploration of this challenging problem. |
| Yichen Zhang (Purdue University) | |
| Online Inference for Robust Policy Evaluation in Reinforcement Learning | |
| Abstract: Reinforcement learning has gained prominence in modern statistics, with policy evaluation being a key component. Unlike traditional machine learning literature on this topic, our work places emphasis on statistical inference for the parameter estimates computed using reinforcement learning algorithms. While most existing analyses assume random rewards to follow standard distributions, limiting their applicability, we embrace the concept of robust statistics in reinforcement learning by simultaneously addressing issues of outlier contamination and heavy-tailed rewards within a unified framework. In this paper, we develop an online robust policy evaluation procedure, and establish the limiting distribution of our estimator, based on its Bahadur representation. Furthermore, we develop a fully-online procedure to efficiently conduct statistical inference based on the asymptotic distribution. This paper bridges the gap between robust statistics and statistical inference in reinforcement learning, offering a more versatile and reliable approach to policy evaluation. |
| Yuhua Zhu (UC Los Angeles) | |
| A PDE-based model-free algorithm for Continuous-time Reinforcement Learning | |
| Abstract: This talk addresses the problem of continuous-time reinforcement learning (RL). When the underlying dynamics remain unknown and only discrete-time observations are available, how can we effectively conduct policy evaluation and policy iteration? We first highlight that while model-free RL algorithms are straightforward to implement, they are often not a reliable approximation of the true value function. On the other hand, model-based PDE approaches are more accurate, but the inverse problem is not easy to solve. To bridge this gap, we introduce a new Bellman equation, PhiBE, which integrates discrete-time information into a PDE formulation. PhiBE allows us to skip the identification of the dynamics and directly evaluate the value function using discrete-time data. Additionally, it offers a more accurate approximation of the true value function, especially in scenarios where the underlying dynamics change slowly. Moreover, we extend PhiBE to higher orders, providing increasingly accurate approximations. |
| José R. Zubizarreta (Harvard University) | |
| An Anatomy of Event Studies: Hypothetical Experiments, Exact Decomposition, and Weighting Diagnostics | |
| Abstract: In recent decades, event studies have emerged as a central methodology in health and social research for evaluating the causal effects of staggered interventions. In this paper, we analyze event studies from experimental design principles for observational studies, with a focus on information borrowing across measurements. We develop robust weighting estimators that increasingly use more information across units and time periods, justified by increasingly stronger assumptions on the treatment assignment and potential outcomes mechanisms. As a particular case of this approach, we offer a novel decomposition of the classical dynamic two-way fixed effects (TWFE) regression estimator for event studies. Our decomposition is expressed in closed form and reveals in finite samples the hypothetical experiment that TWFE regression adjustments approximate. This decomposition offers insights into how standard regression estimators borrow information across different units and times, clarifying and supplementing the notion of forbidden comparison noted in the literature. The proposed approach enables the generalization of treatment effect estimates to a target population and offers new diagnostics for event studies, including covariate balance, sign reversal, effective sample size, and the contribution of each observation to the analysis. We also provide visualization tools for event studies and illustrate them in a case study of the impact of divorce reforms on female suicide. |
| Weinan Wang (LinkedIn) | |
| Real-World Challenges in Ranking, Recommendation, and Causal Inference | |
| Abstract: In this presentation I'll talk about some of the practical challenges we face in the area of experimentation, observational causal inference, ranking and recommendation from my experience at Snapchat and LinkedIn. In addition, I will share some of my personal journey as well, open discussions are welcomed. |
| Arkajyoti Saha (UC Irvine) | |
| Random Forests for Geospatial Data | |
| Abstract: Due to recent advancements in geographical information systems, remote sensing technology, and affordable sensors, we are now faced with datasets that require us to consider spatial dependencies. These geospatial data are often analyzed using the linear mixed model framework, which includes a linear fixed covariate effect and a Gaussian Process (GP)-distributed spatial random effect. However, the assumption that covariate effects are linear is quite limiting. Non-linear modeling of spatial data is gaining popularity, and contemporary extensions of Random Forests (RF) for spatial data diverge from the mixed model setup, giving up inference on the fixed effects and other benefits of utilizing GP. By explicitly modeling the spatial random effects with a GP, we offer a unique and well-principled extension of RF for estimating nonlinear covariate effects in spatial mixed models. Our method extends RF in the same way generalized least square extends ordinary least squares to accommodate for dependence in linear models. For both estimation and prediction with spatial data, our method significantly outperforms classical RF in an extensive simulation study. |
| Abbass Sharif (AXS) | |
| Boosting Ticket Sales and Revenue with Demand-based Dynamic Pricing for Live Music and Sports Events | |
| Abstract: This session will explore the role of demand-based dynamic pricing in optimizing ticket sales for live music and sports events. We’ll examine key factors that drive ticket demand and demonstrate how this approach supports venues and promoters in maximizing revenue while improving attendance. Additionally, we’ll discuss how dynamic pricing ensures fans access tickets at fair market value. Join us to gain practical insights into leveraging dynamic pricing to enhance the live event experience. |
| Letian Yang (University of Southern California) | |
| Algebraic and Statistical Properties of the Partially Regularized Ordinary Least Squares Interpolator | |
| Abstract: Modern deep learning has revealed a surprising statistical phenomenon known as benign overfitting, with high-dimensional linear regression being a prominent example. This paper contributes to ongoing research on the ordinary least squares (OLS) interpolator, focusing on the partial regression setting, where only a subset of coefficients is implicitly regularized. On the algebraic front, we extend Cochran's formula and the leave-one-out residual formula for the partial regularization framework. On the stochastic front, we leverage our algebraic results to design several homoskedastic variance estimators under the Gauss-Markov model. These estimators serve as a basis for conducting statistical inference, albeit with slight conservatism in their performance. Through simulations, we study the finite-sample properties of these variance estimators across various generative models. |
| Boxin Zhao (University of Chicago) | |
| Trans-Glasso: A Transfer Learning Approach to Precision Matrix Estimation | |
| Abstract: Precision matrix estimation is essential in various fields, yet it is challenging when samples for the target study are limited. Transfer learning can enhance estimation accuracy by leveraging data from related source studies. We propose Trans-Glasso, a two-step transfer learning method for precision matrix estimation. First, we obtain initial estimators using a multi-task learning objective that captures shared and unique features across studies. Then, we refine these estimators through differential network estimation to adjust for structural differences between the target and source precision matrices. Under the assumption that most entries of the target precision matrix are shared with source matrices, we derive non-asymptotic error bounds and show that Trans-Glasso achieves minimax optimality under certain conditions. Extensive simulations demonstrate Trans Glasso's superior performance compared to baseline methods, particularly in small-sample settings. We further validate Trans-Glasso in applications to gene networks across brain tissues and protein networks for various cancer subtypes, showcasing its effectiveness in biological contexts. Additionally, we derive the minimax optimal rate for differential network estimation, representing the first such guarantee in this area. |
| Yanfei Zhou (University of Southern California) | |
| Conformal Classification with Equalized Coverage for Adaptively Selected Groups | |
| Abstract: This work introduces a conformal inference method to evaluate uncertainty in classification by generating prediction sets with valid coverage conditional on adaptively chosen features. These features are carefully selected to reflect potential model limitations or biases. This can be useful to find a practical compromise between efficiency -- by providing informative predictions -- and algorithmic fairness -- by ensuring equalized coverage for the most sensitive groups. We demonstrate the validity and effectiveness of this method on simulated and real data sets. |
| John Cherian (Stanford University) | |
| Night-of election modeling via conformal inference | |
| Abstract: We consider a high-stakes application of statistical inference: uncertainty quantification for election night modeling. The Washington Post observes vote counts from early-reporting jurisdictions, e.g., precincts on the East Coast of the United States, and fits a model to these results that predicts the final outcome in each contested race. Quantifying the error of this prediction is crucial; an overconfident prediction can mislead the public and harm the news provider’s reputation. Over the last four years, we have worked on methods to extend conformal prediction to this setting. Working with election data poses many challenges. For example, the data-generating distribution shifts over time, and spatiotemporal correlation can invalidate standard approaches. Variants of our model have been featured in The Washington Post’s coverage of the 2020 and 2022 national United States elections, and the model introduced in this poster was used by the paper in the 2024 presidential election. This is joint work with Lenny Bronner (The Washington Post) and Emmanuel Candès (Stanford). |
| Anav Sood (Stanford University) | |
| Selective inference is easier with p-values | |
| Abstract: Selective inference is a subfield of statistics that enables valid inference after selection of a data-dependent question. In this paper, we introduce selectively dominant p-values, a class of p-values that allow practitioners to easily perform inference after arbitrary selection procedures. Unlike a traditional p-value, whose distribution must stochastically dominate the uniform distribution under the null, a selectively dominant p-value must have a post-selection distribution that stochastically dominates that of a uniform having undergone the same selection process; moreover, this property must hold simultaneously for all possible selection processes. Despite the strength of this condition, we show that all commonly used p-values (e.g., p-values from two-sided testing in parametric families, one-sided testing in monotone likelihood ratio and exponential families, F-tests for linear regression, and permutation tests) are selectively dominant. By recasting two canonical selective inference problems-inference on winners and rank verification-in our selective dominance framework, we provide simpler derivations, a deeper conceptual understanding, and new generalizations and variations of these methods. Additionally, we use our insights to introduce selective variants of methods that combine p-values, such as Fisher's combination test. |
| Tianmin Xie (University of Southern California) | |
| Conformal Classification with Unknown Label Spaces | |
| Abstract: We develop a conformal inference method for classifying labels when the label space is unknown, focusing on species sampling models and conformal p-values. This method constructs prediction sets for new data points when the label space $\mathcal{Y}$ is not fully known or observable in advance. For the hypothesis testing problem of whether a new label is a new species, we introduce a conformal Good-Turing p-value which efficiently tests whether a new observation belongs to a previously unseen species. Moreover, we leverage on features to improve the prediction of new labels. Additionally, the paper proves the optimality of the proposed conformal Good-Turing p-value within the framework of our model. Through numerical experiments, we show how our approach can be applied to complex data, providing a robust framework for predictive modeling in environments where traditional methods fall short. |