In benmarwick/karon: Compositional Data Analysis of Archaeological Artefacts

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

predictSource

The package predictSource provides functions to verify that data can be used to separate sources of samples, to predict the sources of additional samples, and to create plots that evaluate the validity of the predictions. Data can be both quantitative and qualitative. A proposed analysis strategy is to use random forests to evaluate whether the data can separate the sources and to identify the most important predictors if there are many, use a classification tree to understand how the data are used to separate sources, use random forests to predict the sources of unknown samples, and then evaluate the validity of the predictions by ploting the first two principal components of the unknowns with the convex hulls of the known sources. The random forest analysis also produces the probabilities of assignment to each source for each sample; this can be helpful in identifying unknowns that are difficult to classify.The package also contains functions for exploratory data analysis (descriptive statistics, 2- and 3-dimensional plots [the latter can be rotated], tests for 1- and 2-dimensional Gaussian distributions [helpful in identifying outliers]) and multivariate analysis (principal components). A detailed vignette provides examples for the use of each function (using obsidian data in the examples) and some background for classification trees, random forests, and checking for Gaussian distributions.

The motivation for the package was predicting the sources of obsidian artifacts. Archaeologists and geochemists usually do this using 2- and 3-dimensional scatterplots. The functions in this package make predictions much faster than can be done with scatterplots; the principal components graphic identifies objects that are or may be misclassified. Archaeological knowledge should also be used in making predictions.

The figure below shows principal components plots using data sets with the composition of five elements from five obsidian sources and the predicted sources of 91 artifacts, with predictions made from scatterplots (see the vignette for information about these data sets). The left-hand plot shows the convex hulls of the first two principal components from the source data. The second plot shows the locations of artifacts that are outside of their respective predicted source convex hulls. That plot clearly identifies one misclassified artifact (predicted to be from source D but inside the convex hull for source C); the remaining artifacts appear to be correctly classified. For these data, the random forests predictions appear to be correct for all of the artifacts.

library(predictSource)
data(ObsidianSources)
data(ObsidianArtifacts)
analyticVars <- c("Rb", "Sr", "Y", "Zr", "Nb")
sources <- unique(ObsidianSources[, "Code"])
pcaEval <-
  ps_pcaEvaluation(
    SourceData = ObsidianSources,
    unknownData = ObsidianArtifacts,
    SourceGroup = "Code",
    unknownGroup = "Code",
    known_sources = sources,
    predicted_sources = sources,
    AnalyticVars = analyticVars,
    ID = "ID",
    plotAllPoints = TRUE,
    plotHullsOutsidePoints = TRUE,
    plotOutsidePoints = TRUE
  )

The figure below is from a random forests analysis of the artifacts. The figure contains box plots of the source assignment probabilities for each artifact, excluding the probabilities of assignment to the predicted source. This plot identifies the artifacts for which assignment is most difficult. Source C is potentially of most concern. The user can create a data frame with information on artifacts that may be most likely to be misclassified. See the vignette for more details.

library(predictSource)
data(ObsidianSources)
data(ObsidianArtifacts)
analyticVars <- c("Rb", "Sr", "Y", "Zr", "Nb")
saveRandomForest <-
  ps_randomForest(
    data = ObsidianSources,
    GroupVar = "Code",
    Groups = "All",
    AnalyticVars = analyticVars,
    NvarUsed = 3,
    plotErrorRate = FALSE,
    plotImportance = FALSE,
    predictSources = TRUE,
    predictData = ObsidianArtifacts,
    plotSourceProbs = TRUE
  )

Installation

You can install predictSource from GitHub with:

# install.packages("devtools")
devtools::install_github("benmarwick/predictSource")

benmarwick/karon documentation built on July 29, 2023, 10:11 a.m.