In statcodes/nproc: Neyman-Pearson (NP) Classification Algorithms and NP Receiver Operating Characteristic (NP-ROC) Curves

---

title: "A demonstration of the nproc package" author: "Yang Feng, Jingyi Jessica Li and Xin Tong" date: "r Sys.Date()" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{nproc demo} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8}

---

We provide a detailed demo of the usage for the \verb+nproc+ package.

Introduction

Installation and Package Loading

Neyman-Pearson Classifier

Neyman-Pearson Receiver Operator Characteristic Band

Introduction

Let $(X, Y)$ be random variables where $X\in \mathcal{X}$ is a vector of $d$ features and $Y \in {0,1}$ represents a binary class label. A data set that contains independent observations ${(x_i, y_i)}{i=1}^{n+m}$ sampled from the joint distribution of $(X, Y)$ are often divided into training data ${(x_i, y_i)}{i=1}^{n}$ and test data ${(x_i, y_i)}_{i=n+1}^{n+m}$.

Based on training data, a \textit{classifier} $\phi(\cdot)$ is a mapping $\phi: \mathcal{X}\rightarrow {0,1}$ that returns the predicted class label given $X$. Classification error occurs when $\phi(X)\neq Y$, and the binary loss is defined as $1(\phi(X)\neq Y)$, where $1(\cdot)$ denotes the indicator function. The risk is defined as $R(\phi)= E[1(\phi(X)\neq Y)]=P(\phi(X)\neq Y)$, which can be expressed as a weighted sum of type I and II errors: $R(\phi)=P(Y=0)R_0(\phi)+P(Y=1)R_1(\phi),$ where $R_0(\phi)=P(\phi(X)\neq Y|Y=0)$ denotes the type I error, and $R_1(\phi)=P(\phi(X)\neq Y|Y=1)$ denotes the type II error.

The classical classification paradigm aims to find an oracle classifier $\phi^$ by minimizing the risk, $\phi^=\arg\min_{\phi}R(\phi)$.

In contrast, the NP classification paradigm aims to mimic the NP oracle classifier ${\phi}^_\alpha$ with respect to a pre-specified type I error upper bound $\alpha$, $\phi^\alpha=\arg\min{\phi:~R_0(\phi)\leq\alpha}R_1(\phi)$, where $\alpha$ reflects users' conservative attitude (priority) towards the type I error. In practice, since we do not know the class distirbution, we do not expect to enforce that the type I error is bounded from above strictly; instead, we wish the type I error is bounded by $\alpha$ with probability at least $1-\delta$, where $\delta\in(0,1)$ is another user-specified hyper-parameter.

The \verb+nproc+ package provides two essential tools for NP classification: i). a function \verb+npc+ that adapts popular classification algorithms, such as linear discriminant analysis, logistic regression, support vector machines and random forrests to the NP paradigm; and ii) a function \verb+nproc+ that constructs NP-ROC bands, an alternative to the popular ROC curves, that allows visualization and comparison of NP classifiers.

Installation and Package Loading

The simplest way to obtain the \verb+nproc+ package is to install it directly from CRAN. Type the following command in R console:

install.packages("nproc", repos = "http://cran.us.r-project.org")

Users may change the repos options depending on their locations and preferences. Other options such as the directories to which we want to install the packages can be altered in the command. For more details, see help(install.packages).

Here the \verb+nproc+ package has been downloaded and installed to the default directories.

Alternatively, users can download the package source at http://cran.r-project.org/web/packages/nproc/index.html and type Unix commands to install it to the desired location.

Then we can load the nproc package:

library(nproc)

Neyman-Pearson Classifier

Here, we demonstrate how to construct a Neyman-Pearson classifier when a user has priority to control the type I error. Note that a user can also control the type II error under specific levels by switching the class labels and applying the npc function.

In the first step, we create a dataset (x,y) from a logistic regression model with 2 features and sample size 1000.

n = 1000
set.seed(0)
x = matrix(rnorm(n*2),n,2)
c = 1+3*x[,1]
y = rbinom(n,1,1/(1+exp(-c)))

A visualization of the two classes.

plot(x[y==1,],col=1,xlim=c(-4,4),xlab='x1',ylab='x2')
points(x[y==0,],col=2,pch=2)
legend("topright",legend=c('Class 1','Class 0'),col=1:2,pch=c(1,2))

Then, we call the npc function to construct Neyman-Pearson classifiers. Suppose we want to use linear discriminant analysis, we set method = "lda". The default type I error upper bound is alpha=0.05, while the alpha value can be changed to any user specified value in $(0,1)$. The tolerance upper bound $\delta$ also has the default value at $0.05$.

fit = npc(x, y, method = "lda", alpha = 0.05)

We can now evaluate the prediction performance of the NP classifier fit on a test set (xtest, ytest) generated as follows.

xtest = matrix(rnorm(n*2),n,2)
ctest = 1+3*xtest[,1]
ytest = rbinom(n,1,1/(1+exp(-ctest)))

We calculate the overall accuracy of the classifier as well as the realized type I error on test data. It is shown that the type I error is smaller than alpha.Strictly speaking, to demonstrate the effectiveness of the fit classifier under the NP paradigm, we should repeat this experiment many times, and show that in $1-\delta$ of these repetitions, type I error (approximated by empirical type I error on a large test data set) is smaller than alpha.

pred = predict(fit,xtest)
fit.score = predict(fit,x)
accuracy = mean(pred$pred.label==ytest)
cat("Overall Accuracy: ",  accuracy,'\n')
ind0 = which(ytest==0)
typeI = mean(pred$pred.label[ind0]!=ytest[ind0]) #type I error on test set
cat('Type I error: ', typeI, '\n')

The base classification method implemented in the npc function includes the following options.

• logistic: Logistic regression. glm function with family = 'binomial'
• penlog: Penalized logistic regression with LASSO penalty. glmnet in glmnet package
• svm: Support Vector Machines. svm in e1071 package
• randomforest: Random Forest. randomForest in randomForest package
• lda: Linear Discriminant Analysis. lda in MASS package
• nb: Naive Bayes. naiveBayes in e1071 package
• ada: Ada-Boost. ada in ada package

Now, we can change the classification method of choice to logistic regression and change alpha to 0.1.

fit = npc(x, y, method = "logistic", alpha = 0.1)
pred = predict(fit,xtest)
accuracy = mean(pred$pred.label == ytest)
cat("Overall Accuracy: ",  accuracy,'\n')
ind0 = which(ytest == 0)
typeI = mean(pred$pred.label[ind0] != ytest[ind0]) #type I error on test set
cat('Type I error: ', typeI, '\n')

To construct each NP classifier, npc function randomly splits class $0$ sample into two equal halves; the first half, together with the class $1$ sample, is used to train the base classification method, and the second half is reserved to find the threshold on the scores. To increase stability, the nproc package provides implementation of ensembled classifiers, which are majority votes of the NP classifiers constructed from different sample splitting. The parameter value split is the number of random splits. In the following codes, we change its number to $11$.

fit = npc(x, y, method = "logistic", alpha = 0.1, split = 11)
pred = predict(fit,xtest)
accuracy = mean(pred$pred.label == ytest)
cat("Overall Accuracy: ",  accuracy,'\n')
ind0 = which(ytest == 0)
typeI = mean(pred$pred.label[ind0] != ytest[ind0]) #type I error on test set
cat('Type I error: ', typeI, '\n')

Let's compare the performance of different methods on our simulated dataset.

methodlist = c("logistic", "penlog", "randomforest", "svm", 
                                        "lda", "nb", "ada")
for(method in methodlist){
fit = npc(x, y, method = method, alpha = 0.05)
pred = predict(fit,xtest)
accuracy = mean(pred$pred.label==ytest)
cat(method, ': Overall Accuracy: ',  accuracy,'\n')
ind0 = which(ytest==0)
typeI = mean(pred$pred.label[ind0]!=ytest[ind0]) #type I error on test set
cat(method, ': Type I error: ', typeI, '\n')
}

Back to Top

Neyman-Pearson Receiver Operator Characteristic (NP-ROC) Band

The function nproc constructs Neyman-Pearson Receiver Operator Characteristic (NP-ROC) band for an NP classifier. When we plot an NP-ROC band, the horizontal axis is the alpha value, and the vertical axis is the type II error conditioning on training data. When we section the band vertically at each alpha value, the lower point on the segment is a high probability lower bound of the (1-type II error) of the NP classifier corresponding to that alpha value, and the the upper point on the segment is a high probability upper bound of the (1-type II error) of the NP classifier corresponding to that alpha value.

In the following demo, we use the same data as in the previous section.

fit = nproc(x, y, method = "nb")
plot(fit)

Note that, to create an NP-ROC band, there is no need to construct an NP classifier using the npc function first.

When we choose linear discriminant analysis as the base method, we do as follows.

fit2 = nproc(x, y, method = "lda")
plot(fit2)

The compare function compares two NP classifiers at different $\alpha$ values (their $\delta$ values have to be the same) by inespecting their NP-ROC bands. The following code displayes the NP-ROC bands generated by logistic regression with the two predictors. The region marked in red is the alpha range in which the second NP classifier dominates the first, and the black region means otherwise. The uncolored region means we cannot claim supriority of one or the other with confidence.

n = 1000
set.seed(0)
x1 = c(rnorm(n), rnorm(n) + 1)
x2 = c(rnorm(n), rnorm(n)*sqrt(6) + 1)
y = c(rep(0,n), rep(1,n))
fit1 = nproc(x1, y, split = 11, method = 'lda')
fit2 = nproc(x2, y, split = 11, method = 'lda')
v = compare(fit1, fit2)
legend('topleft',legend = c(expression(X[1]),expression(X[2])),col = 1:2,lty = c(1,1))

Lastly, we have implemented a web-based interface using the shiny app. It can be invoked by the following command after the package is loaded. {eval = FALSE} shiny::runApp(system.file('nproc', package='nproc'))

Back to Top

statcodes/nproc documentation built on May 30, 2019, 9:42 a.m.