Asymptotic estimate of depth-based classifiers within the location shift model

TitleAsymptotic estimate of depth-based classifiers within the location shift model
Publication TypeJournal Article
Year of Publication2015
AuthorsGalkin, OA
Abbreviated Key TitleDopov. Nac. akad. nauk Ukr.
DOI10.15407/dopovidi2015.11.030
Issue11
SectionInformation Science and Cybernetics
Pagination30-35
Date Published11/2015
LanguageUkrainian
Abstract

The asymptotic behavior of non-parametric simplicial depth, half-space depth, and spatial depth classifiers is studied under appropriate regularity conditions. The research is carried out for the construction of a maximum depth classifier, when all a priori probabilities of all the competing classes are equal, and the location shift model holds. The constructed maximum depth classifier does not depend on the special parametric form of the dividing surface and classifies the data item to a class, with respect to which the element has a maximum depth of location. The case of unequal a priori probabilities is studied, when different data sets may not belong to the common family of elliptical distributions.

KeywordsBayesian classifier, depth function, Mahalanobis distance
References: 
  1. Zuo Y., Serfling R. The Annals of Statistics, 2000, 28: 484–497.
  2. Serfling R. A depth function and a scale curve based on spatial depth. Statistics and Data Analysis based on L1-Norm and Related Methods, Boston: Birkh¨auser, 2002: 27–36.
  3. Hoeffding W. J. of the American Statistical Association, 1963, 58: 14–27. https://doi.org/10.1080/01621459.1963.10500830
  4. Pollard D.Convergence of Stochastic Processes, New York: Springer, 1984: 1–10. https://doi.org/10.1007/978-1-4612-5254-2_1
  5. Holmes C. C., Adams N. M. J. of the Royal Statistical Society, 2002, 64: 297–304.
  6. Silverman B. W.Density estimation for Statistics and Data Analysis, London: Chapman and Hall, 1986: 1–7. https://doi.org/10.1007/978-1-4899-3324-9_1
  7. Jornsten R., Vardi Y., Zhang C. H. Statistical Data Analysis, 2002: 354–365.