easyHardClassifier: Two-stage Classification Using Easy-to-collect Data Set and...

In ClassifyR: A framework for cross-validated classification problems, with applications to differential variability and differential distribution testing

Description

An alternative implementation to the previously published easy-hard classifier that doesn't do nested cross-validation for speed. In the first stage, each numeric variable is split on all possible midpoints between consecutive ordered values and the samples below the split and above the split are checked to see if they mostly belong to one class. Categorical varaibles are tabulated on factor levels and the count of samples in each class is determined. If any partitions of samples are pure for a class, based on a purity threshold, prediction rules are created. The samples not classified by any rule or classified to two or more classes the same number of times are left to be trained by the hard classifier.

Usage

  ## S4 method for signature 'MultiAssayExperiment'
easyHardClassifierTrain(measurements, easyDatasetID = "clinical", hardDatasetID = names(measurements)[1],
         featureSets = NULL, metaFeatures = NULL, minimumOverlapPercent = 80,
         datasetName = NULL, classificationName = "Easy-Hard Classifier",
         easyClassifierParams = list(minCardinality = 10, minPurity = 0.9),
         hardClassifierParams = list(SelectParams(), TrainParams(), PredictParams()), 
         verbose = 3)
  ## S4 method for signature 'EasyHardClassifier,MultiAssayExperiment'
easyHardClassifierPredict(model, test, predictParams, verbose = 3)

Arguments

measurements

A MultiAssayExperiment object containing the data set The sample classes must be in a column of the DataFrame accessed by colData named "class"

.

easyDatasetID	The name of a data set in measurements or "clinical" to indicate the patient information in the column data be used.
hardDatasetID	The name of a data set in measurements different to the value of easyDatasetID to be used for classifying the samples not classified by the easy classifier.
featureSets	An object of type FeatureSetCollection which defines sets of features or sets of edges.
metaFeatures	Either NULL or a DataFrame which has meta-features of the numeric data of interest.
minimumOverlapPercent	If featureSets stores sets of features, the minimum overlap of feature IDs with measurements for a feature set to be retained in the analysis. If featureSets stores sets of network edges, the minimum percentage of edges with both vertex IDs found in measurements that a set has to have to be retained in the analysis.
datasetName	A name associated with the data set used.
classificationName	A name associated with the classification.
easyClassifierParams	A list of length 2 with names "minCardinality" and "minPurity". The first parameter specifies what the minimum number of samples after a split has to be and the second specifies the minimum proportion of samples in a partition belonging to a particular class.
hardClassifierParams	A list of objects defining the classification to do on the samples which were not predicted by the easy classifier Objects must of of class TransformParams, SelectParams, TrainParams or PredictParams.
model	A trained EasyHardClassifier object.
test	A MultiAssayExperiment object containing the test data.
predictParams	An object of class PredictParams. It specifies the classifier used to make the hard predictions.
verbose	Default: 3. A number between 0 and 3 for the amount of progress messages to give. This function only prints progress messages if the value is 3.

Details

The easy classifier may be NULL if there are no rules that predicted the sample well using the easy data set. The hard classifier may be NULL if all of the samples could be predicted with rules generated using the easy data set or it will simply be a character if all or almost all of the remaining samples belong to one class.

Value

For EasyHardClassifierTrain, the trained two-stage classifier. For EasyHardClassifierPredict, a factor vector of predicted classes.

Author(s)

Dario Strbenac

References

Inspired by: Stepwise Classification of Cancer Samples Using Clinical and Molecular Data, Askar Obulkasim, Gerrit Meijer and Mark van de Wiel 2011, BMC Bioinformatics, Volume 12 article 422, https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-12-422.

Examples

  genesMatrix <- matrix(c(rnorm(90, 9, 1),
                          9.5, 9.4, 5.2, 5.3, 5.4, 9.4, 9.6, 9.9, 9.1, 9.8),
		        ncol = 10, byrow = TRUE)
  colnames(genesMatrix) <- paste("Sample", 1:10)
  rownames(genesMatrix) <- paste("Gene", 1:10)
  genders <- factor(c("Male", "Male", "Female", "Female", "Female",
                      "Female", "Female", "Female", "Female", "Female"))

  # Scenario: Male gender can predict the hard-to-classify Sample 1 and Sample 2.
  clinical <- DataFrame(age = c(31, 34, 32, 39, 33, 38, 34, 37, 35, 36),
                        gender = genders,
                        class = factor(rep(c("Poor", "Good"), each = 5)),
		        row.names = colnames(genesMatrix))
  dataset <- MultiAssayExperiment(ExperimentList(RNA = genesMatrix), clinical)
  selParams <- SelectParams(featureSelection = differentMeansSelection, selectionName = "Difference in Means",
                            resubstituteParams = ResubstituteParams(1:10, "balanced error", "lower"))
  trained <- easyHardClassifierTrain(dataset, easyClassifierParams = list(minCardinality = 2, minPurity = 0.9),
                                     hardClassifierParams = list(selParams, TrainParams(), PredictParams()))

  predictions <- easyHardClassifierPredict(trained, dataset, PredictParams())

ClassifyR documentation built on Nov. 8, 2020, 6:53 p.m.