Search

We establish stability of random forests under the mild condition that the squared response ($Y^2$) does not have a heavy tail. In particular, our analysis holds for the practical version of random forests that is implemented in popular packages like \texttt{randomForest} in \texttt{R}. Empirical results show that stability may persist even beyond our assumption and hold for heavy-tailed $Y^2$. Using the stability property, we prove a non-asymptotic lower bound for the coverage probability of prediction intervals constructed from the out-of-bag error of random forests. With another mild condition that is typically satisfied when $Y$ is continuous, we also establish a complementary upper bound, which can be similarly established for the jackknife prediction interval constructed from an arbitrary stable algorithm. We also discuss the asymptotic coverage probability under assumptions weaker than those considered in previous literature. Our work implies that random forests, with its stability property, is an effective machine learning method that can provide not only satisfactory point prediction but also justified interval prediction at almost no extra computational cost., Comment: NeurIPS 2023

The COVID-19 pandemic was responsible for the cancellation of both the men's and women's 2020 National Collegiate Athletic Association (NCAA) Division 1 basketball tournaments. Starting from the point at which the Division 1 tournaments and any unfinished conference tournaments were cancelled, we deliver closed-form probabilities for each team of making the Division 1 tournaments, had they not been cancelled, aided by use of conformal predictive distributions. We also deliver probabilities of a team winning March Madness, given a tournament bracket. We then compare single-game win probabilities generated with conformal predictive distributions, aptly named conformal win probabilities, to those generated through linear and logistic regression on seven years of historical college basketball data, specifically from the 2014-2015 season through the 2020-2021 season. Conformal win probabilities are shown to be better calibrated than other methods, resulting in more accurate win probability estimates, while requiring fewer distributional assumptions., Comment: preprint submitted to Journal of Quantitative Analysis in Sports, 28 pages without figures; figures included at end of document

Neural networks are among the most powerful nonlinear models used to address supervised learning problems. Similar to most machine learning algorithms, neural networks produce point predictions and do not provide any prediction interval which includes an unobserved response value with a specified probability. In this paper, we proposed the $k$-fold prediction interval method to construct prediction intervals for neural networks based on $k$-fold cross validation. Simulation studies and analysis of 10 real datasets are used to compare the finite-sample properties of the prediction intervals produced by the proposed method and the split conformal (SC) method. The results suggest that the proposed method tends to produce narrower prediction intervals compared to the SC method while maintaining the same coverage probability. Our experimental results also reveal that the proposed $k$-fold prediction interval method produces effective prediction intervals and is especially advantageous relative to competing approaches when the number of training observations is limited., Comment: 8 pages, 4 figures

This paper investigates the problem of adjusting for spatial effects in genomic prediction. Despite being seldomly considered in genomic prediction, spatial effects often affect phenotypic measurements of plants. We consider a Gaussian random field model with an additive covariance structure that incorporates genotype effects, spatial effects and subpopulation effects. An empirical study shows the existence of spatial effects and heterogeneity across different subpopulation families, while simulations illustrate the improvement in selecting genotypically superior plants by adjusting for spatial effects in genomic prediction., Comment: 22 pages, 6 figures, 10 tables

Random forest (RF) methodology is one of the most popular machine learning techniques for prediction problems. In this article, we discuss some cases where random forests may suffer and propose a novel generalized RF method, namely regression-enhanced random forests (RERFs), that can improve on RFs by borrowing the strength of penalized parametric regression. The algorithm for constructing RERFs and selecting its tuning parameters is described. Both simulation study and real data examples show that RERFs have better predictive performance than RFs in important situations often encountered in practice. Moreover, RERFs may incorporate known relationships between the response and the predictors, and may give reliable predictions in extrapolation problems where predictions are required at points out of the domain of the training dataset. Strategies analogous to those described here can be used to improve other machine learning methods via combination with penalized parametric regression techniques., Comment: 12 pages, 5 figures

The additive partially linear model (APLM) combines the flexibility of nonparametric regression with the parsimony of regression models, and has been widely used as a popular tool in multivariate nonparametric regression to alleviate the "curse of dimensionality". A natural question raised in practice is the choice of structure in the nonparametric part, that is, whether the continuous covariates enter into the model in linear or nonparametric form. In this paper, we present a comprehensive framework for simultaneous sparse model identification and learning for ultra-high-dimensional APLMs where both the linear and nonparametric components are possibly larger than the sample size. We propose a fast and efficient two-stage procedure. In the first stage, we decompose the nonparametric functions into a linear part and a nonlinear part. The nonlinear functions are approximated by constant spline bases, and a triple penalization procedure is proposed to select nonzero components using adaptive group LASSO. In the second stage, we refit data with selected covariates using higher order polynomial splines, and apply spline-backfitted local-linear smoothing to obtain asymptotic normality for the estimators. The procedure is shown to be consistent for model structure identification. It can identify zero, linear, and nonlinear components correctly and efficiently. Inference can be made on both linear coefficients and nonparametric functions. We conduct simulation studies to evaluate the performance of the method and apply the proposed method to a dataset on the Shoot Apical Meristem (SAM) of maize genotypes for illustration.

Variable selection in high-dimensional scenarios is of great interested in statistics. One application involves identifying differentially expressed genes in genomic analysis. Existing methods for addressing this problem have some limits or disadvantages. In this paper, we propose distance based variable importance measures to deal with these problems, which is inspired by the Multi-Response Permutation Procedure (MRPP). The proposed variable importance assessments can effectively measure the importance of an individual dimension by quantifying its influence on the differences between multivariate distributions. A backward selection algorithm is developed that can be used in high-dimensional variable selection to discover important variables. Both simulations and real data applications demonstrate that our proposed method enjoys good properties and has advantages over other methods.

When analyzing microarray data, hierarchical models are often used to share information across genes when estimating means and variances or identifying differential expression. Many methods utilize some form of the two-level hierarchical model structure suggested by Kendziorski et al. [Stat. Med. (2003) 22 3899-3914] in which the first level describes the distribution of latent mean expression levels among genes and among differentially expressed treatments within a gene. The second level describes the conditional distribution, given a latent mean, of repeated observations for a single gene and treatment. Many of these models, including those used in Kendziorski's et al. [Stat. Med. (2003) 22 3899-3914] EBarrays package, assume that expression level changes due to treatment effects have the same distribution as expression level changes from gene to gene. We present empirical evidence that this assumption is often inadequate and propose three-level hierarchical models as extensions to the two-level log-normal based EBarrays models to address this inadequacy. We demonstrate that use of our three-level models dramatically changes analysis results for a variety of microarray data sets and verify the validity and improved performance of our suggested method in a series of simulation studies. We also illustrate the importance of accounting for the uncertainty of gene-specific error variance estimates when using hierarchical models to identify differentially expressed genes., Comment: Published in at http://dx.doi.org/10.1214/12-AOAS535 the Annals of Applied Statistics (http://www.imstat.org/aoas/) by the Institute of Mathematical Statistics (http://www.imstat.org)