Journal of Machine Learning Research最新文献

Gaussian processes are widely employed as versatile modelling and predictive tools in spatial statistics, functional data analysis, computer modelling and diverse applications of machine learning. They have been widely studied over Euclidean spaces, where they are specified using covariance functions or covariograms for modelling complex dependencies. There is a growing literature on Gaussian processes over Riemannian manifolds in order to develop richer and more flexible inferential frameworks for non-Euclidean data. While numerical approximations through graph representations have been well studied for the Matérn covariogram and heat kernel, the behaviour of asymptotic inference on the parameters of the covariogram has received relatively scant attention. We focus on asymptotic behaviour for Gaussian processes constructed over compact Riemannian manifolds. Building upon a recently introduced Matérn covariogram on a compact Riemannian manifold, we employ formal notions and conditions for the equivalence of two Matérn Gaussian random measures on compact manifolds to derive the parameter that is identifiable, also known as the microergodic parameter, and formally establish the consistency of the maximum likelihood estimate and the asymptotic optimality of the best linear unbiased predictor. The circle is studied as a specific example of compact Riemannian manifolds with numerical experiments to illustrate and corroborate the theory.

Mixed Membership Models (MMMs) are a popular family of latent structure models for complex multivariate data. Instead of forcing each subject to belong to a single cluster, MMMs incorporate a vector of subject-specific weights characterizing partial membership across clusters. With this flexibility come challenges in uniquely identifying, estimating, and interpreting the parameters. In this article, we propose a new class of Dimension-Grouped MMMs ( $\text{Gro-}{\text{M}}^3\text{s}$ ) for multivariate categorical data, which improve parsimony and interpretability. In $\text{Gro-}{\text{M}}^3\text{s}$ , observed variables are partitioned into groups such that the latent membership is constant for variables within a group but can differ across groups. Traditional latent class models are obtained when all variables are in one group, while traditional MMMs are obtained when each variable is in its own group. The new model corresponds to a novel decomposition of probability tensors. Theoretically, we derive transparent identifiability conditions for both the unknown grouping structure and model parameters in general settings. Methodologically, we propose a Bayesian approach for Dirichlet $\text{Gro-}{\text{M}}^3\text{s}$ to inferring the variable grouping structure and estimating model parameters. Simulation results demonstrate good computational performance and empirically confirm the identifiability results. We illustrate the new methodology through applications to a functional disability survey dataset and a personality test dataset.

Insights into complex, high-dimensional data can be obtained by discovering features of the data that match or do not match a model of interest. To formalize this task, we introduce the "data selection" problem: finding a lower-dimensional statistic-such as a subset of variables-that is well fit by a given parametric model of interest. A fully Bayesian approach to data selection would be to parametrically model the value of the statistic, nonparametrically model the remaining "background" components of the data, and perform standard Bayesian model selection for the choice of statistic. However, fitting a nonparametric model to high-dimensional data tends to be highly inefficient, statistically and computationally. We propose a novel score for performing data selection, the "Stein volume criterion (SVC)", that does not require fitting a nonparametric model. The SVC takes the form of a generalized marginal likelihood with a kernelized Stein discrepancy in place of the Kullback-Leibler divergence. We prove that the SVC is consistent for data selection, and establish consistency and asymptotic normality of the corresponding generalized posterior on parameters. We apply the SVC to the analysis of single-cell RNA sequencing data sets using probabilistic principal components analysis and a spin glass model of gene regulation.

Reinforcement learning (RL) has shown great promise in estimating dynamic treatment regimes which take into account patient heterogeneity. However, health-outcome information, used as the reward for RL methods, is often not well coded but rather embedded in clinical notes. Extracting precise outcome information is a resource-intensive task, so most of the available well-annotated cohorts are small. To address this issue, we propose a semi-supervised learning (SSL) approach that efficiently leverages a small-sized labeled data set with actual outcomes observed and a large unlabeled data set with outcome surrogates. In particular, we propose a semi-supervised, efficient approach to Q-learning and doubly robust off-policy value estimation. Generalizing SSL to dynamic treatment regimes brings interesting challenges: 1) Feature distribution for Q-learning is unknown as it includes previous outcomes. 2) The surrogate variables we leverage in the modified SSL framework are predictive of the outcome but not informative of the optimal policy or value function. We provide theoretical results for our Q function and value function estimators to understand the degree of efficiency gained from SSL. Our method is at least as efficient as the supervised approach, and robust to bias from mis-specification of the imputation models.

Zero-inflated count data arise in a wide range of scientific areas such as social science, biology, and genomics. Very few causal discovery approaches can adequately account for excessive zeros as well as various features of multivariate count data such as overdispersion. In this paper, we propose a new zero-inflated generalized hypergeometric directed acyclic graph (ZiG-DAG) model for inference of causal structure from purely observational zero-inflated count data. The proposed ZiG-DAGs exploit a broad family of generalized hypergeometric probability distributions and are useful for modeling various types of zero-inflated count data with great flexibility. In addition, ZiG-DAGs allow for both linear and nonlinear causal relationships. We prove that the causal structure is identifiable for the proposed ZiG-DAGs via a general proof technique for count data, which is applicable beyond the proposed model for investigating causal identifiability. Score-based algorithms are developed for causal structure learning. Extensive synthetic experiments as well as a real dataset with known ground truth demonstrate the superior performance of the proposed method against state-of-the-art alternative methods in discovering causal structure from observational zero-inflated count data. An application of reverse-engineering a gene regulatory network from a single-cell RNA-sequencing dataset illustrates the utility of ZiG-DAGs in practice.

Multiple testing is a commonly used tool in modern data science. Sometimes, the hypotheses are embedded in a space; the distances between the hypotheses reflect their co-null/co-alternative patterns. Properly incorporating the distance information in testing will boost testing power. Hence, we developed a new multiple testing framework named Distance Assisted Recursive Testing (DART). DART features in joint artificial intelligence (AI) and statistics modeling. It has two stages. The first stage uses AI models to construct an aggregation tree that reflects the distance information. The second stage uses statistical models to embed the testing on the tree and control the false discovery rate. Theoretical analysis and numerical experiments demonstrated that DART generates valid, robust, and powerful results. We applied DART to a clinical trial in the allogeneic stem cell transplantation study to identify the gut microbiota whose abundance was impacted by post-transplant care.

Statistical inference of directed relations given some unspecified interventions (i.e., the intervention targets are unknown) is challenging. In this article, we test hypothesized directed relations with unspecified interventions. First, we derive conditions to yield an identifiable model. Unlike classical inference, testing directed relations requires to identify the ancestors and relevant interventions of hypothesis-specific primary variables. To this end, we propose a peeling algorithm based on nodewise regressions to establish a topological order of primary variables. Moreover, we prove that the peeling algorithm yields a consistent estimator in low-order polynomial time. Second, we propose a likelihood ratio test integrated with a data perturbation scheme to account for the uncertainty of identifying ancestors and interventions. Also, we show that the distribution of a data perturbation test statistic converges to the target distribution. Numerical examples demonstrate the utility and effectiveness of the proposed methods, including an application to infer gene regulatory networks. The R implementation is available at https://github.com/chunlinli/intdag.

As machine learning powered decision-making becomes increasingly important in our daily lives, it is imperative to strive for fairness in the underlying data processing. We propose a pre-processing algorithm for fair data representation via which $L^2\left(\text{ℙ}\right)$ -objective supervised learning results in estimations of the Pareto frontier between prediction error and statistical disparity. Particularly, the present work applies the optimal affine transport to approach the post-processing Wasserstein barycenter characterization of the optimal fair $L^2$ -objective supervised learning via a pre-processing data deformation. Furthermore, we show that the Wasserstein geodesics from learning outcome marginals to their barycenter characterizes the Pareto frontier between $L^2$ -loss and total Wasserstein distance among the marginals. Numerical simulations underscore the advantages: (1) the pre-processing step is compositive with arbitrary conditional expectation estimation supervised learning methods and unseen data; (2) the fair representation protects the sensitive information by limiting the inference capability of the remaining data with respect to the sensitive data; (3) the optimal affine maps are computationally efficient even for high-dimensional data.

Risk modeling with electronic health records (EHR) data is challenging due to no direct observations of the disease outcome and the high-dimensional predictors. In this paper, we develop a surrogate assisted semi-supervised learning approach, leveraging small labeled data with annotated outcomes and extensive unlabeled data of outcome surrogates and high-dimensional predictors. We propose to impute the unobserved outcomes by constructing a sparse imputation model with outcome surrogates and high-dimensional predictors. We further conduct a one-step bias correction to enable interval estimation for the risk prediction. Our inference procedure is valid even if both the imputation and risk prediction models are misspecified. Our novel way of ultilizing unlabelled data enables the high-dimensional statistical inference for the challenging setting with a dense risk prediction model. We present an extensive simulation study to demonstrate the superiority of our approach compared to existing supervised methods. We apply the method to genetic risk prediction of type-2 diabetes mellitus using an EHR biobank cohort.

A widely recognized difficulty in federated learning arises from the statistical heterogeneity among clients: local datasets often originate from distinct yet not entirely unrelated probability distributions, and personalization is, therefore, necessary to achieve optimal results from each individual's perspective. In this paper, we show how the excess risks of personalized federated learning using a smooth, strongly convex loss depend on data heterogeneity from a minimax point of view, with a focus on the FedAvg algorithm (McMahan et al., 2017) and pure local training (i.e., clients solve empirical risk minimization problems on their local datasets without any communication). Our main result reveals an approximate alternative between these two baseline algorithms for federated learning: the former algorithm is minimax rate optimal over a collection of instances when data heterogeneity is small, whereas the latter is minimax rate optimal when data heterogeneity is large, and the threshold is sharp up to a constant. As an implication, our results show that from a worst-case point of view, a dichotomous strategy that makes a choice between the two baseline algorithms is rate-optimal. Another implication is that the popular FedAvg following by local fine tuning strategy is also minimax optimal under additional regularity conditions. Our analysis relies on a new notion of algorithmic stability that takes into account the nature of federated learning.