Cross validation for estimating the classification rate.
dirda.cv(x, ina, folds = NULL, nfolds = 10, stratified = FALSE,
type = c("vmf", "iag", "esag", "kent", "sc", "pkbd", "purka"),
seed = NULL, B = 1000)
A list including:
A vector with the estimated performance of each classifier.
The bootstrap bias corrected performance.
A matrix with the data in Eulcidean coordinates, i.e. unit vectors. The matrix must have three columns, only spherical data are currently supported.
A variable indicating the groupings.
Do you already have a list with the folds? If not, leave this NULL.
How many folds to create?
Should the folds be created in a stratified way? i.e. keeping the distribution of the groups similar through all folds?
If seed is TRUE, the results will always be the same.
The type of classifier to use. The avaliable options are "vmf" (von Mises-Fisher distribution), "iag" (IAG distribution), "esag" (ESAG distribution), "kent" (Kent distribution), "sc" and "sc2" (spherical Cauchy distribution), "pkbd" and "pkbd2" (Poisson kernel based distribution), and "purka" (Purkayastha distribution). The difference between "sc" and "sc2" and between "pkbd" and "pkbd2" is that the first uses the Newton-Raphson algorithm and it is faster, whereas the second uses a hybrid algorithm that does not require the Hessian matrix, but in large dimensions the second will be faster. You can chose any of them or all of them. Note that "kent" works only with spherical data.
If you used k-NN, should a bootstrap correction of the bias be applied? If yes, 1000 is a good value.
Michail Tsagris.
R implementation and documentation: Michail Tsagris mtsagris@uoc.gr.
Cross-validation for the estimation of the performance of a classifier.
The estimated performance of the best classifier is overestimated. After the cross-valdiation procedure, the predicted values produced by all classifiers are colelcted, from all folds, in an \(n \times M\) matrix, where \(n\) is the number of samples and \(M\) is the number of all classifiers used. We sample rows (predictions) with replacement from P and denote them as the in-sample values. The non re-sampled rows are denoted as out-of-sample values. The performance of each classifier in the insample rows is calculated and the classifier with the optimal performance is selected, followed by the calculation of performance in the out-of-sample values. This process is repeated B times and the average performance is returned. The only computational overhead is with the repetitive resampling and calculation of the performance, i.e. no model or classifier is fitted nor trained. For more information see Tsamardinos et al. (2018).
The good thing with the function is that you can run any method you want by supplying the
folds yourselves using the command makefolds
. Then suppose you want to run another method. By suppying the same folds you will be able to have comparative results for all methods.
Tsagris M. (2024). Directional data analysis using the spherical Cauchy and the Poisson-kernel based distribution. https://arxiv.org/pdf/2409.03292
Tsagris M. and Alenazi A. (2019). Comparison of discriminant analysis methods on the sphere. Communications in Statistics: Case Studies, Data Analysis and Applications, 5(4), 467--491.
Mardia, K. V. and Jupp, P. E. (2000). Directional statistics. Chicester: John Wiley & Sons.
Morris J. E. and Laycock P. J. (1974). Discriminant analysis of directional data. Biometrika, 61(2): 335--341.
Tsamardinos I., Greasidou E. and Borboudakis G. (2018). Machince Learning, 107(12): 1895--1922.
dirda, dirknn, knn.reg
x <- rvmf(300, rnorm(3), 10)
ina <- sample.int(2, 300, replace = TRUE)
dirda.cv(x, ina, B = 1)
Run the code above in your browser using DataLab