On Maximum Depth and Related Classifiers

Abstract.  Over the last couple of decades, data depth has emerged as a powerful exploratory and inferential tool for multivariate data analysis with wide-spread applications. This paper investigates the possible use of different notions of data depth in non-parametric discriminant analysis. First, we consider the situation where the prior probabilities of the competing populations are all equal and investigate classifiers that assign an observation to the population with respect to which it has the maximum location depth. We propose a different depth-based classification technique for unequal prior problems, which is also useful for equal prior cases, especially when the populations have different scatters and shapes. We use some simulated data sets as well as some benchmark real examples to evaluate the performance of these depth-based classifiers. Large sample behaviour of the misclassification rates of these depth-based non-parametric classifiers have been derived under appropriate regularity conditions.     

Problems with Procrustes analysis

Procrustes analysis is applied to three different data sets. From this, it is inferred that the uncritical use of this technique can be highly misleading. It is suggested that an examination of the internal structure of the data is needed to clarify the results and one form of such examination is used as an illustration in the text.     

Hard or Soft Classification? Large-margin Unified Machines

Margin-based classifiers have been popular in both machine learning and statistics for classification problems. Among numerous classifiers, some are hard classifiers while some are soft ones. Soft classifiers explicitly estimate the class conditional probabilities and then perform classification based on estimated probabilities. In contrast, hard classifiers directly target on the classification decision boundary without producing the probability estimation. These two types of classifiers are based on different philosophies and each has its own merits. In this paper, we propose a novel family of large-margin classifiers, namely large-margin unified machines (LUMs), which covers a broad range of margin-based classifiers including both hard and soft ones. By offering a natural bridge from soft to hard classification, the LUM provides a unified algorithm to fit various classifiers and hence a convenient platform to compare hard and soft classification. Both theoretical consistency and numerical performance of LUMs are explored. Our numerical study sheds some light on the choice between hard and soft classifiers in various classification problems.     

Kernel density classification and boosting: an L2 analysis

Kernel density estimation is a commonly used approach to classification. However, most of the theoretical results for kernel methods apply to estimation per se and not necessarily to classification. In this paper we show that when estimating the difference between two densities, the optimal smoothing parameters are increasing functions of the sample size of the complementary group, and we provide a small simluation study which examines the relative performance of kernel density methods when the final goal is classification.A relative newcomer to the classification portfolio is boosting, and this paper proposes an algorithm for boosting kernel density classifiers. We note that boosting is closely linked to a previously proposed method of bias reduction in kernel density estimation and indicate how it will enjoy similar properties for classification. We show that boosting kernel classifiers reduces the bias whilst only slightly increasing the variance, with an overall reduction in error. Numerical examples and simulations are used to illustrate the findings, and we also suggest further areas of research.     

Robust Model-Free Multiclass Probability Estimation

Classical statistical approaches for multiclass probability estimation are typically based on regression techniques such as multiple logistic regression, or density estimation approaches such as linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA). These methods often make certain assumptions on the form of probability functions or on the underlying distributions of subclasses. In this article, we develop a model-free procedure to estimate multiclass probabilities based on large-margin classifiers. In particular, the new estimation scheme is employed by solving a series of weighted large-margin classifiers and then systematically extracting the probability information from these multiple classification rules. A main advantage of the proposed probability estimation technique is that it does not impose any strong parametric assumption on the underlying distribution and can be applied for a wide range of large-margin classification methods. A general computational algorithm is developed for class probability estimation. Furthermore, we establish asymptotic consistency of the probability estimates. Both simulated and real data examples are presented to illustrate competitive performance of the new approach and compare it with several other existing methods.     

Consistency of Procrustes Estimators

Lele has shown that the Procrustes estimator of form is inconsistent and raised the question about the consistency of the Procrustes estimator of shape. In this paper the consistency of estimators of form and shape is studied under various assumptions. In particular, it is shown that the Procrustes estimator of shape is consistent under the assumption of an isotropic error distribution and that consistency breaks down if the assumption of isotropy is relaxed. The relevance of these results for practical shape analysis is discussed. As a by-product, some new results are derived for the offset uniform distribution from directional data.     

Regularisation,interpolation and visualisation of diffusion tensor images using non-Euclidean statistics

Practical statistical analysis of diffusion tensor images is considered, and we focus primarily on methods that use metrics based on Euclidean distances between powers of diffusion tensors. First, we describe a family of anisotropy measures based on a scale invariant power-Euclidean metric, which are useful for visualisation. Some properties of the measures are derived and practical considerations are discussed, with some examples. Second, we discuss weighted Procrustes methods for diffusion tensor imaging interpolation and smoothing, and we compare methods based on different metrics on a set of examples as well as analytically. We establish a key relationship between the principal-square-root-Euclidean metric and the size-and-shape Procrustes metric on the space of symmetric positive semi-definite tensors. We explain, both analytically and by experiments, why the size-and-shape Procrustes metric may be preferred in practical tasks of interpolation, extrapolation and smoothing, especially when observed tensors are degenerate or when a moderate degree of tensor swelling is desirable. Third, we introduce regularisation methodology, which is demonstrated to be useful for highlighting features of prior interest and potentially for segmentation. Finally, we compare several metrics in a data set of human brain diffusion-weighted magnetic resonance imaging, and point out similarities between several of the non-Euclidean metrics but important differences with the commonly used Euclidean metric.     

Regression Methods for Combining Multiple Classifiers

As no single classification method outperforms other classification methods under all circumstances, decision-makers may solve a classification problem using several classification methods and examine their performance for classification purposes in the learning set. Based on this performance, better classification methods might be adopted and poor methods might be avoided. However, which single classification method is the best to predict the classification of new observations is still not clear, especially when some methods offer similar classification performance in the learning set. In this article we present various regression and classical methods, which combine several classification methods to predict the classification of new observations. The quality of the combined classifiers is examined on some real data. Nonparametric regression is the best method of combining classifiers.     

Partial order relations for classification comparisons

The Bayes classification rule offers the optimal classifier, minimizing the classification error rate, whereas the Neyman–Pearson lemma offers the optimal family of classifiers to maximize the detection rate for any given false alarm rate. These motivate studies on comparing classifiers based on similarities between the classifiers and the optimal. In this article, we define partial order relations on classifiers and families of classifiers, based on rankings of rate function values and rankings of test function values, respectively. Each partial order relation provides a sufficient condition, which yields better classification error rates or better performance on the receiver operating characteristic analysis. Various examples and applications of the partial order theorems are discussed to provide comparisons of classifiers and families of classifiers, including the comparison of cross-validation methods, training data that contains outliers, and labelling errors in training data. The Canadian Journal of Statistics 48: 152–166; 2020 © 2019 Statistical Society of Canada     

Have I seen you before? Principles of Bayesian predictive classification revisited

A general inductive Bayesian classification framework is considered using a simultaneous predictive distribution for test items. We introduce a principle of generative supervised and semi-supervised classification based on marginalizing the joint posterior distribution of labels for all test items. The simultaneous and marginalized classifiers arise under different loss functions, while both acknowledge jointly all uncertainty about the labels of test items and the generating probability measures of the classes. We illustrate for data from multiple finite alphabets that such classifiers achieve higher correct classification rates than a standard marginal predictive classifier which labels all test items independently, when training data are sparse. In the supervised case for multiple finite alphabets the simultaneous and the marginal classifiers are proven to become equal under generalized exchangeability when the amount of training data increases. Hence, the marginal classifier can be interpreted as an asymptotic approximation to the simultaneous classifier for finite sets of training data. It is also shown that such convergence is not guaranteed in the semi-supervised setting, where the marginal classifier does not provide a consistent approximation.     

A classification updating procedure motivated by high-content screening data

The current paradigm for the identification of candidate drugs within the pharmaceutical industry typically involves the use of high-throughput screens. High-content screening (HCS) is the term given to the process of using an imaging platform to screen large numbers of compounds for some desirable biological activity. Classification methods have important applications in HCS experiments, where they are used to predict which compounds have the potential to be developed into new drugs. In this paper, a new classification method is proposed for batches of compounds where the rule is updated sequentially using information from the classification of previous batches. This methodology accounts for the possibility that the training data are not a representative sample of the test data and that the underlying group distributions may change as new compounds are analysed. This technique is illustrated on an example data set using linear discriminant analysis, k-nearest neighbour and random forest classifiers. Random forests are shown to be superior to the other classifiers and are further improved by the additional updating algorithm in terms of an increase in the number of true positives as well as a decrease in the number of false positives.     

Non-crossing large-margin probability estimation and its application to robust SVM via preconditioning

Many large-margin classifiers such as the Support Vector Machine (SVM) sidestep estimating conditional class probabilities and target the discovery of classification boundaries directly. However, estimation of conditional class probabilities can be useful in many applications. Wang, Shen, and Liu (2008) bridged the gap by providing an interval estimator of the conditional class probability via bracketing. The interval estimator was achieved by applying different weights to positive and negative classes and training the corresponding weighted large-margin classifiers. They propose to estimate the weighted large-margin classifiers individually. However, empirically the individually estimated classification boundaries may suffer from crossing each other even though, theoretically, they should not.In this work, we propose a technique to ensure non-crossing of the estimated classification boundaries. Furthermore, we take advantage of the estimated conditional class probabilities to precondition our training data. The standard SVM is then applied to the preconditioned training data to achieve robustness. Simulations and real data are used to illustrate their finite sample performance.     

A comparison of the post selection error rate behaviour of the normal linear and quadratic discriminant rules

The normal linear discriminant rule (NLDR) and the normal quadratic discriminant rule (NQDR) are popular classifiers when working with normal populations. Several papers in the literature have been devoted to a comparison of these rules with respect to classification performance. An aspect which has, however, not received any attention is the effect of an initial variable selection step on the relative performance of these classification rules. Cross model validation variabie selection has been found to perform well in the linear case, and can be extended to the quadratic case. We report the results of a simulation study comparing the NLDR and the NQDR with respect to the post variable selection classification performance. It is of interest that the NQDR generally benefits from an initial variable selection step. We also comment briefly on the problem of estimating the post selection error rates of the two rules.     

Tilting methods for assessing the influence of components in a classifier

Summary.  Many contemporary classifiers are constructed to provide good performance for very high dimensional data. However, an issue that is at least as important as good classification is determining which of the many potential variables provide key information for good decisions. Responding to this issue can help us to determine which aspects of the datagenerating mechanism (e.g. which genes in a genomic study) are of greatest importance in terms of distinguishing between populations. We introduce tilting methods for addressing this problem. We apply weights to the components of data vectors, rather than to the data vectors themselves (as is commonly the case in related work). In addition we tilt in a way that is governed by L ₂-distance between weight vectors, rather than by the more commonly used Kullback–Leibler distance. It is shown that this approach, together with the added constraint that the weights should be non-negative, produces an algorithm which eliminates vector components that have little influence on the classification decision. In particular, use of the L ₂-distance in this problem produces properties that are reminiscent of those that arise when L ₁-penalties are employed to eliminate explanatory variables in very high dimensional prediction problems, e.g. those involving the lasso. We introduce techniques that can be implemented very rapidly, and we show how to use bootstrap methods to assess the accuracy of our variable ranking and variable elimination procedures.     

Ideal bootstrap estimation of expected prediction error for k-nearest neighbor classifiers: Applications for classification and error assessment

Euclidean distance k-nearest neighbor (k-NN) classifiers are simple nonparametric classification rules. Bootstrap methods, widely used for estimating the expected prediction error of classification rules, are motivated by the objective of calculating the ideal bootstrap estimate of expected prediction error. In practice, bootstrap methods use Monte Carlo resampling to estimate the ideal bootstrap estimate because exact calculation is generally intractable. In this article, we present analytical formulae for exact calculation of the ideal bootstrap estimate of expected prediction error for k-NN classifiers and propose a new weighted k-NN classifier based on resampling ideas. The resampling-weighted k-NN classifier replaces the k-NN posterior probability estimates by their expectations under resampling and predicts an unclassified covariate as belonging to the group with the largest resampling expectation. A simulation study and an application involving remotely sensed data show that the resampling-weighted k-NN classifier compares favorably to unweighted and distance-weighted k-NN classifiers.     

Finite-sample analysis of impacts of unlabeled data and their labeling mechanisms in linear discriminant analysis

It is widely believed that unlabeled data are promising for improving prediction accuracy in classification problems. Although theoretical studies about when/how unlabeled data are beneficial exist, an actual prediction improvement has not been sufficiently investigated for a finite sample in a systematic manner. We investigate the impact of unlabeled data in linear discriminant analysis and compare the error rates of the classifiers estimated with/without unlabeled data. Our focus is a labeling mechanism that characterizes the probabilistic structure of occurrence of labeled cases. Results imply that an extremely small proportion of unlabeled data has a large effect on the analysis results.     

Classification using semiparametric mixtures

A new density-based classification method that uses semiparametric mixtures is proposed. Like other density-based classifiers, it first estimates the probability density function for the observations in each class, with a semiparametric mixture, and then classifies a new observation by the highest posterior probability. By making a proper use of a multivariate nonparametric density estimator that has been developed recently, it is able to produce adaptively smooth and complicated decision boundaries in a high-dimensional space and can thus work well in such cases. Issues specific to classification are studied and discussed. Numerical studies using simulated and real-world data show that the new classifier performs very well as compared with other commonly used classification methods.     

Applying classification algorithms in practice

In this paper we present a perspective on the overall process of developing classifiers for real-world classification problems. Specifically, we identify, categorize and discuss the various problem-specific factors that influence the development process. Illustrative examples are provided to demonstrate the iterative nature of the process of applying classification algorithms in practice. In addition, we present a case study of a large scale classification application using the process framework described, providing an end-to-end example of the iterative nature of the application process. The paper concludes that the process of developing classification applications for operational use involves many factors not normally considered in the typical discussion of classification models and algorithms.     

Bayesian Procrustes analysis with applications to hydrology

In this paper, we introduce Procrustes analysis in a Bayesian framework, by treating the classic Procrustes regression equation from a Bayesian perspective, while modeling shapes in two dimensions. The Bayesian approach allows us to compute point estimates and credible sets for the full Procrustes fit parameters. The methods are illustrated through an application to radar data from short-term weather forecasts (nowcasts), a very important problem in hydrology and meteorology.     

An ensemble method for concept drift in nonstationary environment

Most statistical and data-mining algorithms assume that data come from a stationary distribution. However, in many real-world classification tasks, data arrive over time and the target concept to be learned from the data stream may change accordingly. Many algorithms have been proposed for learning drifting concepts. To deal with the problem of learning when the distribution generating the data changes over time, dynamic weighted majority was proposed as an ensemble method for concept drift. Unfortunately, this technique considers neither the age of the classifiers in the ensemble nor their past correct classification. In this paper, we propose a method that takes into account expert's age as well as its contribution to the global algorithm's accuracy. We evaluate the effectiveness of our proposed method by using m classifiers and training a collection of n-fold partitioning of the data. Experimental results on a benchmark data set show that our method outperforms existing ones.