In frankp-0/fastMCD: Robust multivariate estimation of location and scatter

knitr::opts_chunk$set(
  collapse = TRUE,
  out.width = "100%"
  )

fastMCD

Overview

fastMCD performs robust mean and covariance estimation using the FAST-MCD algorithm described by Rousseeuw and van Driessen (1999).

Introduction

The minimum covariance determinant (MCD) method obtains robust estimates of location and scale for multivariate data by identifying a subset of h observations with the lowest determinant in its covariance matrix. This method is highly robust and accurate but is infeasible for large datasets. The FAST-MCD algorithm, described by Rousseeuw and van Driessen (1999), approximates the MCD estimate with an iterative procedure applied to a set of initial subsamples. FAST-MCD is far faster than MCD and typically is accurate even for large datasets. Superior implementations exist in R already, however the fastMCD package represents my attempt to independently implement FAST-MCD and understand the underlying algorithm.

Installation

devtools::install_github("frankp-0/fastMCD", build_vignettes = T)

Note that installation may take > 1 minute if vignettes are built.

Usage

The following examples demonstrate the usage of fastMCD for small (350 observations) and large (5000 observations) datasets.

smallX <- MASS::mvrnorm(n = 350, mu = rep(0, 3), Sigma = diag(rep(1, 3)))
fastMCD(smallX)

bigX <- MASS::mvrnorm(n = 5000, mu = rep(0, 3), Sigma = diag(rep(1, 3)))
fastMCD(bigX)

Users may optionally specify their own argument for h = # of observations to subsample.

X <- MASS::mvrnorm(n = 350, mu = rep(0, 3), Sigma = diag(rep(1, 3)))
fastMCD(bigX, h = 500)

frankp-0/fastMCD documentation built on Dec. 20, 2021, 8:51 a.m.