Article
Automation & Control Systems
Asad Haris, Noah Simon, Ali Shojaie
Summary: In this paper, we propose a unified framework for estimating and analyzing high-dimensional generalized additive models. The framework defines a large class of penalized regression estimators and presents an efficient computational algorithm. We prove min-max optimal convergence bounds for this class and characterize the rate of convergence when a compatibility condition is not met. Additionally, we show the link between the optimal penalty parameters for structure and sparsity penalties in our framework.
JOURNAL OF MACHINE LEARNING RESEARCH
(2022)
Article
Computer Science, Interdisciplinary Applications
Marinela Capanu, Mihai Giurcanu, Colin B. Begg, Mithat Gonen
Summary: Introduces a novel variable selection method named OPT-STABS for low-dimensional generalized linear models. OPT-STABS repeatedly subsamples the data, minimizes AIC over a sequence of nested models for each subsample, and includes predictors selected in the minimum AIC model in a large fraction of the subsamples. It outperforms other methods in most settings and exhibits competitive performance in the rest.
COMPUTATIONAL STATISTICS & DATA ANALYSIS
(2023)
Article
Neurosciences
Oystein Sorensen, Andreas M. Brandmaier, Didac Macia, Klaus Ebmeier, Paolo Ghisletta, Rogier A. Kievit, Athanasia M. Mowinckel, Kristine B. Walhovd, Rene Westerhausen, Anders Fjell
Summary: Analyzing data from multiple neuroimaging studies has the potential to increase statistical power and systematically investigate differences between studies. Meta-analytic methods, such as meta-GAM, provide a way to combine aggregated quantities and overcome challenges associated with sharing neuroimaging datasets. This approach is especially beneficial in lifespan neuroscience and imaging genetics.
Article
Computer Science, Theory & Methods
Xinmin Li, Haozhe Liang, Wolfgang Haerdle, Hua Liang
Summary: We propose test statistics based on the penalized spline to decide between generalized linear models and generalized partially linear models. The numerical performance of the proposed statistics is comparable to that of their kernel-based competitors, which have been shown to be asymptotically normal in the literature (Hardle et al. in J Am Stat Assoc 93:1461-1474, 1998). We also numerically explore the possibility of using the proposed statistics for goodness of fit checking for GLM. The proposed procedures are illustrated to analyze two datasets.
STATISTICS AND COMPUTING
(2023)
Article
Mathematics
Zuleyka Diaz Martinez, Jose Fernandez Menendez, Luis Javier Garcia Villalba
Summary: This study compares the commonly used GLM and GAM approaches in automobile insurance pricing. The results show that GAMs are a powerful alternative to GLMs in terms of prediction quality, complexity of use, and execution time.
Article
Operations Research & Management Science
Michel Grabisch, Christophe Labreuche, Mustapha Ridaoui
Summary: This paper examines the decomposition problem of generalized additive independence (GAI) models, particularly 2-additive GAI models, and introduces the concept of well-formed decomposition. In the case of discrete variables, the paper provides a practical procedure to obtain a monotone and well-formed decomposition.
ANNALS OF OPERATIONS RESEARCH
(2022)
Article
Computer Science, Artificial Intelligence
Bin Gu, Chenkang Zhang, Zhouyuan Huo, Heng Huang
Summary: This paper proposes a new doubly stochastic optimization algorithm (DSGAM) for solving generalized additive models (GAM). The algorithm can scale up additive models in both sample size and dimensionality, and has been proven to have a fast convergence rate. Experimental results on large-scale benchmark datasets demonstrate the fast convergence and significant reduction in computational time compared with existing algorithms, while maintaining similar generalization performance.
Article
Automation & Control Systems
Jing Ouyang, Kean Ming Tan, Gongjun Xu
Summary: This paper proposes a debiasing approach for high-dimensional regression models with hidden confounding. By adjusting for the effects induced by the unmeasured confounders, the proposed method addresses the issue of invalidity in standard debiasing methods. The consistency and asymptotic normality of the proposed debiased estimator are established, and its finite sample performance is demonstrated through numerical studies and a genetic data set.
JOURNAL OF MACHINE LEARNING RESEARCH
(2023)
Article
Neurosciences
Paul A. Thompson, Kate E. Watkins, Zoe V. J. Woodhead, Dorothy V. M. Bishop
Summary: This study investigates how to align the analysis of brain lateralization using fTCD data with the statistical methods commonly used in fMRI. The results show that using complex GAM method has the lowest measurement error and can more accurately identify cases of bilateral language. Additionally, the GAM-based approach can efficiently analyze more complex designs that include interactions between tasks.
HUMAN BRAIN MAPPING
(2023)
Article
Statistics & Probability
Ye Tian, Yang Feng
Summary: In this work, the authors study the transfer learning problem under high-dimensional generalized linear models (GLMs). They propose a transfer learning algorithm on GLM and a detection approach to detect informative sources. The effectiveness of the algorithms is verified through extensive simulations and a real-data experiment.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2022)
Article
Computer Science, Interdisciplinary Applications
Benjamin D. Youngman
Summary: This article introduces the R package evgam, which provides functions for fitting extreme value distributions including the generalized extreme value and generalized Pareto distributions. The package also supports quantile regression using the asymmetric Laplace distribution, which is useful for estimating high thresholds. The main addition of package evgam is the ability to model extreme value distribution parameters using generalized additive models with objectively estimated smoothness using Laplace's method.
JOURNAL OF STATISTICAL SOFTWARE
(2022)
Article
Mechanics
Marco Mondelli, Ramji Venkataramanan
Summary: The study focuses on estimating signals from measurements obtained via a generalized linear model, proposing an AMP algorithm initialized with a spectral estimator and rigorously characterizing its performance in the high-dimensional limit.
JOURNAL OF STATISTICAL MECHANICS-THEORY AND EXPERIMENT
(2022)
Article
Computer Science, Interdisciplinary Applications
Javier Roca-Pardinas, Maria Xose Rodriguez-Alvarez, Stefan Sperlich
Summary: A package is introduced for a large family of popular semiparametric regression models, known as generalized structured models. The kernel-based weighted smooth backfitting estimator is statistically efficient and well-understood in terms of its asymptotic properties. It allows for flexible handling of various data characteristics and provides convenient data processing and presentation options.
Article
Environmental Sciences
Ganesh Kumar Rajahmundry, Chandrasekhar Garlapati, Ponnusamy Senthil Kumar, Ratna Surya Alwi, Dai-Viet N. Vo
Summary: In this study, adsorption data for different solutes on activated carbon were considered and modeled using twelve commonly known isotherm models. The Langmuir-Freundlich isotherm was found to be the most effective in correlating the adsorption data among the models compared.
Article
Economics
Xiaodong Song, Saralees Nadarajah
Summary: Two new models were proposed for the popular extramarital affairs data set due to Fair, showing better fits than many known models. The fits were evaluated using AIC, residual plots, comparison of observed and predicted counts, and formal tests.