Nearest shrunken centroid (NSC) classifiers for high-dimensional,
low-sample-size data such as gene expression. Implements the original
homoscedastic NSC with soft thresholding (Prediction Analysis of
Microarrays; Tibshirani et al. 2002
Heteroscedastic nearest shrunken centroid (NSC) classification for high-dimensional, low-sample-size data such as gene expression.
The package implements six classifiers:
| Label | Thresholding | Variance in test statistic and discriminant |
|---|---|---|
| STh | soft | pooled: the original PAM (Tibshirani et al. 2002) |
| HTh | hard | pooled (Sahtout, Wang and Ghimire 2024) |
| OTh | order | pooled (Sahtout, Wang and Ghimire 2024) |
| STh3 | soft | class-specific (Sahtout and Wang) |
| HTh3 | hard | class-specific (Sahtout and Wang) |
| OTh3 | order | class-specific (Sahtout and Wang) |
The threshold is chosen by the refined ("deep") cross-validation search of Sahtout, Wang and Ghimire (2024).
# from a local copy of the source package
install.packages("hetNSC_0.1.0.tar.gz", repos = NULL, type = "source")
# or, once it is on GitHub
# remotes::install_github("<user>/hetNSC")
The package only depends on base R and MASS.
Expression matrices have genes in rows and samples in columns (as in pamr).
library(hetNSC)
set.seed(8900)
dat <- simulate_hetero_data(ntr1 = 15, ntr2 = 20, nte1 = 50, nte2 = 50,
p = 1000, a = 14)
fit <- hnsc(dat$xtr, dat$ytr, method = "HTh3") # or method = "hard", variance = "hetero"
fit
pred <- predict(fit, dat$xte) # class labels
mean(pred != dat$yte) # test error
predict(fit, dat$xte, type = "posterior") # posterior probabilities
fit$selected.genes # genes used
| Function | Purpose |
|---|---|
hnsc() |
Fit a classifier with the threshold selected by deep CV |
predict() |
Class labels, posterior probabilities, shrunken centroids or selected genes |
hnsc_train(), hnsc_cv(), hnsc_select() |
Threshold grid fit, cross-validation, deep CV search |
compare_nsc() |
Repeated-run comparison of homoscedastic vs heteroscedastic versions (protocol of the paper); summary() and recall_table() summarise it |
simulate_hetero_data() |
Correlated-gene simulation designs (class B only, or both classes) |
compute_rho2() |
Relative heterogeneity and bias-variance crossover check (Section 2.4) |
make_cm(), per_class_recall(), balanced_acc() |
Class-level metrics |
hnsc_methods() |
Table of the six classifiers |
See vignette("hetNSC") for a walk-through, and
system.file("scripts", "simulation_correlated.R", package = "hetNSC") for the
full correlated-gene simulation studies.
The original research scripts are kept in inst/legacy. For the same random
seed the package returns the same selected thresholds, cross-validation
errors, test errors, numbers of genes and posterior probabilities as those
scripts; tests/testthat/test-legacy.R checks this.
citation("hetNSC")
GPL-2 (GNU General Public License, version 2 only; see LICENSE). This is the licence of the R package pamr by Hastie, Tibshirani, Narasimhan and Chu, from which parts of the code are adapted.