Feng Liang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Feng Liang has a total of 56 co-authors. Here are the top co-authors by number of shared publications:
- Christopher Qian
University of Illinois Urbana-Champaign
2 shared publications - Daniel Khashabi
University of Illinois Urbana-Champaign Urbana 61801 IL
2 shared publications - Jason Adams
Sandia National Laboratories (SNL-NM), Albuquerque, NM (United States)
2 shared publications - Jin Wang
Department of Statistics University of Illinois at Urbana Champaign
2 shared publications Anlei Dong
Yahoo! Labs, Sunnyvale, CA, USA
1 shared publication- Anwesha Chakravarti
Department of Statistics University of Illinois at Urbana Champaign United States
1 shared publication
Latest Publications
- Open Access
- 3 April 2025
- Published by Taylor & Francis in Technometrics
- Vol. 67 (3) , 481-492
Regression Recalibration by Learning PIT Map Values
- 3 April 2025
- Published by Taylor & Francis in Technometrics
- Vol. 67 (3) , 481-492
Open AccessRegression Recalibration by Learning PIT Map Values
- Open Access
Bayesian Smoothing and Feature Selection Using variational Automatic Relevance Determination
This study introduces Variational Automatic Relevance Determination (VARD), a novel approach tailored for fitting sparse additive regression models in high-dimensional settings. VARD distinguishes itself by its ability to independently assess the smoothness of each feature while enabling precise determination of whether a feature's contribution to the response is zero, linear, or nonlinear. Further, an efficient coordinate descent algorithm is introduced to implement VARD. Empirical evaluations on simulated and real-world data underscore VARD's superiority over alternative variable selection methods for additive models.
Open AccessBayesian Smoothing and Feature Selection Using variational Automatic Relevance Determination
This study introduces Variational Automatic Relevance Determination (VARD), a novel approach tailored for fitting sparse additive regression models in high-dimensional settings. VARD distinguishes itself by its ability to independently assess the smoothness of each feature while enabling precise determination of whether a feature's contribution to the response is zero, linear, or nonlinear. Further, an efficient coordinate descent algorithm is introduced to implement VARD. Empirical evaluations on simulated and real-world data underscore VARD's superiority over alternative variable selection methods for additive models.
- Open Access
Bayesian Variable Selection and Sparse Estimation for High-Dimensional Graphical Models
We introduce a novel Bayesian approach for both covariate selection and sparse precision matrix estimation in the context of high-dimensional Gaussian graphical models involving multiple responses. Our approach provides a sparse estimation of the three distinct sparsity structures: the regression coefficient matrix, the conditional dependency structure among responses, and between responses and covariates. This contrasts with existing methods, which typically focus on any two of these structures but seldom achieve simultaneous sparse estimation for all three. A key aspect of our method is that it leverages the structural sparsity information gained from the presence of irrelevant covariates in the dataset to introduce covariate-level sparsity in the precision and regression coefficient matrices. This is achieved through a Bayesian conditional random field model using a hierarchical spike and slab prior setup. Despite the non-convex nature of the problem, we establish statistical accuracy for points in the high posterior density region, including the maximum-a-posteriori (MAP) estimator. We also present an efficient Expectation-Maximization (EM) algorithm for computing the estimators. Through simulation experiments, we demonstrate the competitive performance of our method, particularly in scenarios with weak signal strength in the precision matrices. Finally, we apply our method to a bike-share dataset, showcasing its predictive performance.
Open AccessBayesian Variable Selection and Sparse Estimation for High-Dimensional Graphical Models
We introduce a novel Bayesian approach for both covariate selection and sparse precision matrix estimation in the context of high-dimensional Gaussian graphical models involving multiple responses. Our approach provides a sparse estimation of the three distinct sparsity structures: the regression coefficient matrix, the conditional dependency structure among responses, and between responses and covariates. This contrasts with existing methods, which typically focus on any two of these structures but seldom achieve simultaneous sparse estimation for all three. A key aspect of our method is that it leverages the structural sparsity information gained from the presence of irrelevant covariates in the dataset to introduce covariate-level sparsity in the precision and regression coefficient matrices. This is achieved through a Bayesian conditional random field model using a hierarchical spike and slab prior setup. Despite the non-convex nature of the problem, we establish statistical accuracy for points in the high posterior density region, including the maximum-a-posteriori (MAP) estimator. We also present an efficient Expectation-Maximization (EM) algorithm for computing the estimators. Through simulation experiments, we demonstrate the competitive performance of our method, particularly in scenarios with weak signal strength in the precision matrices. Finally, we apply our method to a bike-share dataset, showcasing its predictive performance.
- Open Access
- 18 September 2024
- Published by Wiley in Statistical Analysis and Data Mining: An ASA Data Science Journal
- Vol. 17 (5)
Quantifying Epistemic Uncertainty in Binary Classification via Accuracy Gain
Recently, a surge of interest has been given to quantifying epistemic uncertainty (EU), the reducible portion of uncertainty due to lack of data. We propose a novel EU estimator in the binary classification setting, as the posterior expected value of the empirical gain in accuracy between the current prediction and the optimal prediction. In order to validate the performance of our EU estimator, we introduce an experimental procedure where we take an existing dataset, remove a set of points, and compare the estimated EU with the observed change in accuracy. Through real and simulated data experiments, we demonstrate the effectiveness of our proposed EU estimator.
- 18 September 2024
- Published by Wiley in Statistical Analysis and Data Mining: An ASA Data Science Journal
- Vol. 17 (5)
Open AccessQuantifying Epistemic Uncertainty in Binary Classification via Accuracy Gain
Recently, a surge of interest has been given to quantifying epistemic uncertainty (EU), the reducible portion of uncertainty due to lack of data. We propose a novel EU estimator in the binary classification setting, as the posterior expected value of the empirical gain in accuracy between the current prediction and the optimal prediction. In order to validate the performance of our EU estimator, we introduce an experimental procedure where we take an existing dataset, remove a set of points, and compare the estimated EU with the observed change in accuracy. Through real and simulated data experiments, we demonstrate the effectiveness of our proposed EU estimator.
- 19 July 2022
- Published by Taylor & Francis in Journal of the American Statistical Association
- Vol. 118 (544) , 2860-2875
Learning Topic Models: Identifiability and Finite-Sample Analysis
Topic models provide a useful text-mining tool for learning, extracting, and discovering latent structures in large text corpora. Although a plethora of methods have been proposed for topic modeling, lacking in the literature is a formal theoretical investigation of the statistical identifiability and accuracy of latent topic estimation. In this paper, we propose a maximum likelihood estimator (MLE) of latent topics based on a specific integrated likelihood that is naturally connected to the concept, in computational geometry, of volume minimization. Our theory introduces a new set of geometric conditions for topic model identifiability, conditions that are weaker than conventional separability conditions, which typically rely on the existence of pure topic documents or of anchor words. Weaker conditions allow a wider and thus potentially more fruitful investigation. We conduct finite-sample error analysis for the proposed estimator and discuss connections between our results and those of previous investigations. We conclude with empirical studies employing both simulated and real datasets.
- 19 July 2022
- Published by Taylor & Francis in Journal of the American Statistical Association
- Vol. 118 (544) , 2860-2875
Learning Topic Models: Identifiability and Finite-Sample Analysis
Topic models provide a useful text-mining tool for learning, extracting, and discovering latent structures in large text corpora. Although a plethora of methods have been proposed for topic modeling, lacking in the literature is a formal theoretical investigation of the statistical identifiability and accuracy of latent topic estimation. In this paper, we propose a maximum likelihood estimator (MLE) of latent topics based on a specific integrated likelihood that is naturally connected to the concept, in computational geometry, of volume minimization. Our theory introduces a new set of geometric conditions for topic model identifiability, conditions that are weaker than conventional separability conditions, which typically rely on the existence of pure topic documents or of anchor words. Weaker conditions allow a wider and thus potentially more fruitful investigation. We conduct finite-sample error analysis for the proposed estimator and discuss connections between our results and those of previous investigations. We conclude with empirical studies employing both simulated and real datasets.
- 3 March 2022
- Published by Springer Nature in Lifetime Data Analysis
- Vol. 28 (2) , 282-318
Bayesian penalized Buckley-James method for high dimensional bivariate censored regression models
For high dimensional gene expression data, one important goal is to identify a small number of genes that are associated with progression of the disease or survival of the patients. In this paper, we consider the problem of variable selection for multivariate survival data. We propose an estimation procedure for high dimensional accelerated failure time (AFT) models with bivariate censored data. The method extends the Buckley-James method by minimizing a penalized \(L_2\) loss function with a penalty function induced from a bivariate spike-and-slab prior specification. In the proposed algorithm, censored observations are imputed using the Kaplan-Meier estimator, which avoids a parametric assumption on the error terms. Our empirical studies demonstrate that the proposed method provides better performance compared to the alternative procedures designed for univariate survival data regardless of whether the true events are correlated or not, and conceptualizes a formal way of handling bivariate survival data for AFT models. Findings from the analysis of a myeloma clinical trial using the proposed method are also presented.- 3 March 2022
- Published by Springer Nature in Lifetime Data Analysis
- Vol. 28 (2) , 282-318
Bayesian penalized Buckley-James method for high dimensional bivariate censored regression models
For high dimensional gene expression data, one important goal is to identify a small number of genes that are associated with progression of the disease or survival of the patients. In this paper, we consider the problem of variable selection for multivariate survival data. We propose an estimation procedure for high dimensional accelerated failure time (AFT) models with bivariate censored data. The method extends the Buckley-James method by minimizing a penalized \(L_2\) loss function with a penalty function induced from a bivariate spike-and-slab prior specification. In the proposed algorithm, censored observations are imputed using the Kaplan-Meier estimator, which avoids a parametric assumption on the error terms. Our empirical studies demonstrate that the proposed method provides better performance compared to the alternative procedures designed for univariate survival data regardless of whether the true events are correlated or not, and conceptualizes a formal way of handling bivariate survival data for AFT models. Findings from the analysis of a myeloma clinical trial using the proposed method are also presented.- 1 January 2022
- Published by Statistica Sinica (Institute of Statistical Science) in Statistica Sinica
- Vol. 32 (1) , 131-152
Bayesian Estimation of Gaussian Conditional Random Fields
- 1 January 2022
- Published by Statistica Sinica (Institute of Statistical Science) in Statistica Sinica
- Vol. 32 (1) , 131-152
Bayesian Estimation of Gaussian Conditional Random Fields
- Top Cited Papers
- 27 May 2021
- Published by Wiley in WIREs Water
- Vol. 8 (5) , e1533
Machine learning for hydrologic sciences: An introductory overview
The hydrologic community has experienced a surge in interest in machine learning in recent years. This interest is primarily driven by rapidly growing hydrologic data repositories, as well as success of machine learning in various academic and commercial applications, now possible due to increasing accessibility to enabling hardware and software. This overview is intended for readers new to the field of machine learning. It provides a non‐technical introduction, placed within a historical context, to commonly used machine learning algorithms and deep learning architectures. Applications in hydrologic sciences are summarized next, with a focus on recent studies. They include the detection of patterns and events such as land use change, approximation of hydrologic variables and processes such as rainfall‐runoff modeling, and mining relationships among variables for identifying controlling factors. The use of machine learning is also discussed in the context of integrated with process‐based modeling for parameterization, surrogate modeling, and bias correction. Finally, the article highlights challenges of extrapolating robustness, physical interpretability, and small sample size in hydrologic applications.This article is categorized under: Science of Water
- 27 May 2021
- Published by Wiley in WIREs Water
- Vol. 8 (5) , e1533
Top Cited PapersMachine learning for hydrologic sciences: An introductory overview
The hydrologic community has experienced a surge in interest in machine learning in recent years. This interest is primarily driven by rapidly growing hydrologic data repositories, as well as success of machine learning in various academic and commercial applications, now possible due to increasing accessibility to enabling hardware and software. This overview is intended for readers new to the field of machine learning. It provides a non‐technical introduction, placed within a historical context, to commonly used machine learning algorithms and deep learning architectures. Applications in hydrologic sciences are summarized next, with a focus on recent studies. They include the detection of patterns and events such as land use change, approximation of hydrologic variables and processes such as rainfall‐runoff modeling, and mining relationships among variables for identifying controlling factors. The use of machine learning is also discussed in the context of integrated with process‐based modeling for parameterization, surrogate modeling, and bias correction. Finally, the article highlights challenges of extrapolating robustness, physical interpretability, and small sample size in hydrologic applications.This article is categorized under: Science of Water
- 26 October 2020
- Published by Wiley in Statistical Analysis and Data Mining: An ASA Data Science Journal
- Vol. 14 (1) , 31-40
Complementary dimension reduction
- 26 October 2020
- Published by Wiley in Statistical Analysis and Data Mining: An ASA Data Science Journal
- Vol. 14 (1) , 31-40
Complementary dimension reduction
- Open Access
- 10 March 2019
- Published by MDPI AG in Sustainability
- Vol. 11 (5) , 1474
Impact of Building Design Parameters on Daylighting Metrics Using an Analysis, Prediction, and Optimization Approach Based on Statistical Learning Technique
Daylighting metrics are used to predict the daylight availability within a building and assess the performance of a fenestration solution. In this process, building design parameters are inseparable from these metrics; therefore, we need to know which parameters are truly important and how they impact performance. The purpose of this study is to explore the relationship between building design attributes and existing daylighting metrics based on a new methodology we are proposing. This methodology involves statistical learning. It is an emerging methodology that helps us to analyze a large quantity of output data and the impact of a large number of design variables. In particular, we can use these statistical methodologies to analyze which features are important, which ones are not, and the type of relationships they have. Using these techniques, statistical models may be created to predict daylighting metric values for different building types and design solutions. In this article we will outline how this methodology works, and analyze the building design features that have the strongest impact on daylighting performance.
- 10 March 2019
- Published by MDPI AG in Sustainability
- Vol. 11 (5) , 1474
Open AccessImpact of Building Design Parameters on Daylighting Metrics Using an Analysis, Prediction, and Optimization Approach Based on Statistical Learning Technique
Daylighting metrics are used to predict the daylight availability within a building and assess the performance of a fenestration solution. In this process, building design parameters are inseparable from these metrics; therefore, we need to know which parameters are truly important and how they impact performance. The purpose of this study is to explore the relationship between building design attributes and existing daylighting metrics based on a new methodology we are proposing. This methodology involves statistical learning. It is an emerging methodology that helps us to analyze a large quantity of output data and the impact of a large number of design variables. In particular, we can use these statistical methodologies to analyze which features are important, which ones are not, and the type of relationships they have. Using these techniques, statistical models may be created to predict daylighting metric values for different building types and design solutions. In this article we will outline how this methodology works, and analyze the building design features that have the strongest impact on daylighting performance.