Search

Hall and Robinson (2009) proposed and analyzed the use of bagged cross-validation to choose the bandwidth of a kernel density estimator. They established that bagging greatly reduces the noise inherent in ordinary cross-validation, and hence leads to a more efficient bandwidth selector. The asymptotic theory of Hall and Robinson (2009) assumes that $N$, the number of bagged subsamples, is $\infty$. We expand upon their theoretical results by allowing $N$ to be finite, as it is in practice. Our results indicate an important difference in the rate of convergence of the bagged cross-validation bandwidth for the cases $N=\infty$ and $N<\infty$. Simulations quantify the improvement in statistical efficiency and computational speed that can result from using bagged cross-validation as opposed to a binned implementation of ordinary cross-validation. The performance of thebagged bandwidth is also illustrated on a real, very large, data set. Finally, a byproduct of our study is the correction of errors appearing in the Hall and Robinson (2009) expression for the asymptotic mean squared error of the bagging selector., Comment: 37 pages, 9 figures

Feature or variable selection is a problem inherent to large data sets. While many methods have been proposed to deal with this problem, some can scale poorly with the number of predictors in a data set. Screening methods scale linearly with the number of predictors by checking each predictor one at a time, and are a tool used to decrease the number of variables to consider before further analysis or variable selection. For classification, there is a variety of techniques. There are parametric based screening tests, such as t-test or SIS based screening, and non-parametric based screening tests, such as Kolmogorov distance based screening, and MV-SIS. We propose a method for variable screening that uses Bayesian-motivated tests, compare it to SIS based screening, and provide example applications of the method on simulated and real data. It is shown that our screening method can lead to improvements in classification rate. This is so even when our method is used in conjunction with a classifier, such as DART, which is designed to select a sparse subset of variables. Finally, we propose a classifier based on kernel density estimates that in some cases can produce dramatic improvements in classification rates relative to DART.

A growing demand for handling uncertainties and risks in performance-driven building design decision-making has challenged conventional design methods. Thus, researchers in this field lean towards viable alternatives to using deterministic design methods, e.g., probabilistic methods. This research addresses the challenges associated with conventional methods of performance-driven building design, i.e., ignoring the existing uncertainties and lacking a systematic framework to incorporate risk assessment in building performance analysis. This research introduces a framework (BIMProbE) to integrate BIM-based parametric tools with building probabilistic performance analysis to facilitate uncertainty analysis and risk assessment in performance-based building design decision-making. A hypothetical building design scenario is used to demonstrate the application of the proposed framework. The results show that the probabilistic method leads to a different performance ranking order than the deterministic method. Also, the probabilistic method allows evaluating design options based on different risk attitudes., Comment: 18 pages, 6 figures, CAADFutures2021 conference, this is a pre-print

This paper provides a framework for estimating the mean and variance of a high-dimensional normal density. The main setting considered is a fixed number of vector following a high-dimensional normal distribution with unknown mean and diagonal covariance matrix. The diagonal covariance matrix can be known or unknown. If the covariance matrix is unknown, the sample size can be as small as $2$. The proposed estimator is based on the idea that the unobserved pairs of mean and variance for each dimension are drawn from an unknown bivariate distribution, which we model as a mixture of normal-inverse gammas. The mixture of normal-inverse gamma distributions provides advantages over more traditional empirical Bayes methods, which are based on a normal-normal model. When fitting a mixture model, we are essentially clustering the unobserved mean and variance pairs for each dimension into different groups, with each group having a different normal-inverse gamma distribution. The proposed estimator of each mean is the posterior mean of shrinkage estimates, each of which shrinks a sample mean towards a different component of the mixture distribution. Similarly, the proposed estimator of variance has an analogous interpretation in terms of sample variances and components of the mixture distribution. If diagonal covariance matrix is known, then the sample size can be as small as $1$, and we treat the pairs of known variance and unknown mean for each dimension as random observations coming from a flexible mixture of normal-inverse gamma distributions.

We present an efficient method to estimate cross-validation bandwidth parameters for kernel density estimation in very large datasets where ordinary cross-validation is rendered highly inefficient, both statistically and computationally. Our approach relies on calculating multiple cross-validation bandwidths on partitions of the data, followed by suitable scaling and averaging to return a partitioned cross-validation bandwidth for the entire dataset. The partitioned cross-validation approach produces substantial computational gains over ordinary cross-validation. We additionally show that partitioned cross-validation can be statistically efficient compared to ordinary cross-validation. We derive analytic expressions for the asymptotically optimal number of partitions and study its finite sample accuracy through a detailed simulation study. We additionally propose a permuted version of partitioned cross-validation which attains even higher efficiency. Theoretical properties of the estimators are studied and the methodology is applied to the Higgs Boson dataset with 11 million observations

Fully robust OSCV is a modification of the OSCV method that produces consistent bandwidth in the cases of smooth and nonsmooth regression functions. The current implementation of the method uses the kernel $H_I$ that is almost indistinguishable from the Gaussian kernel on the interval $[-4,4]$, but has negative tails. The theoretical properties and practical performances of the $H_I$- and $\phi$-based OSCV versions are compared. The kernel $H_I$ tends to produce too low bandwidths in the smooth case. The $H_I$-based OSCV curves are shown to have wiggles appearing in the neighborhood of zero. The kernel $H_I$ uncovers sensitivity of the OSCV method to a tiny modification of the kernel used for the cross-validation purposes. The recently found robust bimodal kernels tend to produce OSCV curves with multiple local minima. The problem of finding a robust unimodal nonnegative kernel remains open., Comment: 9 figures, 2 tables

A new method of bandwidth selection for kernel density estimators is proposed. The method, termed indirect cross-validation, or ICV, makes use of so-called selection kernels. Least squares cross-validation (LSCV) is used to select the bandwidth of a selection-kernel estimator, and this bandwidth is appropriately rescaled for use in a Gaussian kernel estimator. The proposed selection kernels are linear combinations of two Gaussian kernels, and need not be unimodal or positive. Theory is developed showing that the relative error of ICV bandwidths can converge to 0 at a rate of $n^{-1/4}$, which is substantially better than the $n^{-1/10}$ rate of LSCV. Interestingly, the selection kernels that are best for purposes of bandwidth selection are very poor if used to actually estimate the density function. This property appears to be part of the larger and well-documented paradox to the effect that "the harder the estimation problem, the better cross-validation performs." The ICV method uniformly outperforms LSCV in a simulation study, a real data example, and a simulated example in which bandwidths are chosen locally., Comment: 26 pages, 10 figures

In this paper we provide insight into the empirical properties of indirect cross-validation (ICV), a new method of bandwidth selection for kernel density estimators. First, we describe the method and report on the theoretical results used to develop a practical-purpose model for certain ICV parameters. Next, we provide a detailed description of a numerical study which shows that the ICV method usually outperforms least squares cross-validation (LSCV) in finite samples. One of the major advantages of ICV is its increased stability compared to LSCV. Two real data examples show the benefit of using both ICV and a local version of ICV., Comment: 22 pages, 21 figures

A rank-based test of the null hypothesis that a regressor has no effect on a response variable is proposed and analyzed. This test is identical in structure to the order selection test but with the raw data replaced by ranks. The test is nonparametric in that it is consistent against virtually any smooth alternative, and is completely distribution free for all sample sizes. The asymptotic distribution of the rank-based order selection statistic is obtained and seen to be the same as that of its raw data counterpart. Exact small sample critical values of the test statistic are provided as well. It is shown that the Pitman-Noether efficiency of the proposed rank test compares very favorably with that of the order selection test. In fact, their asymptotic relative efficiency is identical to that of the Wilcoxon signed rank and $t$-tests. An example involving microarray data illustrates the usefulness of the rank test in practice., Comment: Published in at http://dx.doi.org/10.1214/193940307000000103 the IMS Collections (http://www.imstat.org/publications/imscollections.htm) by the Institute of Mathematical Statistics (http://www.imstat.org)

We propose and analyze nonparametric tests of the null hypothesis that a function belongs to a specified parametric family. The tests are based on BIC approximations, \pi_{BIC}, to the posterior probability of the null model, and may be carried out in either Bayesian or frequentist fashion. We obtain results on the asymptotic distribution of \pi_{BIC} under both the null hypothesis and local alternatives. One version of \pi_{BIC}, call it \pi_{BIC}^*, uses a class of models that are orthogonal to each other and growing in number without bound as sample size, n, tends to infinity. We show that \sqrtn(1-\pi_{BIC}^*) converges in distribution to a stable law under the null hypothesis. We also show that \pi_{BIC}^* can detect local alternatives converging to the null at the rate \sqrt\log n/n. A particularly interesting finding is that the power of the \pi_{BIC}^*-based test is asymptotically equal to that of a test based on the maximum of alternative log-likelihoods. Simulation results and an example involving variable star data illustrate desirable features of the proposed tests., Comment: Published at http://dx.doi.org/10.1214/009053604000000805 in the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)