assemble<-,Asset-method {Allspice}

Finish Asset contents

Description

Trains a new classification model.

Usage

assemble(obj) <- value

Arguments

obj

An object of the class Asset.

value

A list that containts training data, see details.

Details

The value argument must contain three named elements: title, dat and bits. Optional predictors and covariates elements can also be included.

The title is a descriptive identifier for the asset that will be displayed by the Classifier object in report().

The dat element is a matrix that contains the training samples. The variables are organized into named rows and the samples into named columns. Non-finite values are not allowed.

The predictors element contains the names of the input variables that should be used for training the model. If empty, all inputs are used for automatic feature selection and subsequent training steps.

The covariates element contains additional information for constructing the final classification models. Unlike the data matrix, variables are organized into named columns and the samples are stored as rows. Non-finite values are not allowed.

The bits element contains labels for category memberships. Three formats are supported: 1) a character vector of named elements that contains non-empty strings, 2) a matrix or a data frame with row names and a single column of non-empty values, and 3) a matrix or a data frame with multiple columns that contain binary values where 1s indicate category membership (the name of the column is the name of the category). Overlapping categories are allowed.

The final asset is assembled in six steps. First, the training data are standardized and normalized. Second, input variables are sorted according to their univariate classification performance. Third, redundant features are excluded by testing the sorted variables for mutual correlations; this produces an optimized listing of non-redundant features that are the most predictive of the category labels. Fourth, mean centroids are calculated for each category. Fifth, training samples are matched to their nearest centroids and the distances collected as preliminary predictor scores. Lastly, logistic regression models are fitted to the preliminary scores, covariates and category labels to enable the calculation of standardized predictor scores for new data.

Value

Updates the Asset object.

Examples

# Prepare training data.
simu <- bcellALL(200)
materials <- list(title="Simutypes")
materials$dat <- simu$counts
materials$covariates <- simu$metadata[,c("MALE","AGE")]
materials$bits <- simu$metadata[,"SUBTYPE",drop=FALSE]

# Assemble classification asset.
bALL <- asset()
assemble(bALL) <- materials

# Export asset into a new folder.
tpath <- tempfile()
export(bALL, folder = tpath)

# Create a classifier.
cls <- classifier(tpath, verbose = FALSE)

# Classify new samples.
simu <- bcellALL(5)
covariates(cls) <- simu$metadata
profiles(cls) <- simu$counts
primary <- predictions(cls)[[1]]
print(primary[,c("LABEL","PROX","EXCL")])