Pre-Processing of Predictors
Pre-processing transformation (centering, scaling etc.) can be estimated from the training data and applied to any data set with the same variables.
preProcess(x, ...)"preProcess"(x, method = c("center", "scale"), thresh = 0.95, pcaComp = NULL, na.remove = TRUE, k = 5, knnSummary = mean, outcome = NULL, fudge = .2, numUnique = 3, verbose = FALSE, ...)"predict"(object, newdata, ...)
- a matrix or data frame. Non-numeric predictors are allowed but will be ignored.
- a character vector specifying the type of processing. Possible values are "BoxCox", "YeoJohnson", "expoTrans", "center", "scale", "range", "knnImpute", "bagImpute", "medianImpute", "pca", "ica", "spatialSign", "zv", "nzv", and "conditionalX" (see Details below)
- a cutoff for the cumulative percent of variance to be retained by PCA
- the specific number of PCA components to keep. If specified, this over-rides
- a logical; should missing values be removed from the calculations?
- an object of class
- a matrix or data frame of new data to be pre-processed
- the number of nearest neighbors from the training set to use for imputation
- function to average the neighbor values per column during imputation
- a numeric or factor vector for the training set outcomes. This can be used to help estimate the Box-Cox transformation of the predictor variables (see Details below)
- a tolerance value: Box-Cox transformation lambda values within +/-fudge will be coerced to 0 and within 1+/-fudge will be coerced to 1.
- how many unique values should
yhave to estimate the Box-Cox transformation?
- a logical: prints a log as the computations proceed
- additional arguments to pass to
fastICA, such as
In all cases, transformations and operations are estimated using the data in
x and these operations are applied to new data using these values; nothing is recomputed when using the
The Box-Cox (
method = "BoxCox"), Yeo-Johnson (
method = "YeoJohnson"), and exponential transformations (
method = "expoTrans")have been "repurposed" here: they are being used to transform the predictor variables. The Box-Cox transformation was developed for transforming the response variable while another method, the Box-Tidwell transformation, was created to estimate transformations of predictor data. However, the Box-Cox method is simpler, more computationally efficient and is equally effective for estimating power transformations. The Yeo-Johnson transformation is similar to the Box-Cox model but can accommodate predictors with zero and/or negative values (while the predictors values for the Box-Cox transformation must be strictly positive.) The exponential transformation of Manly (1976) can also be used for positive or negative data.
method = "center" subtracts the mean of the predictor's data (again from the data in
x) from the predictor values while
method = "scale" divides by the standard deviation.
The "range" transformation scales the data to be within [0, 1]. If new samples have values larger or smaller than those in the training set, values will be outside of this range.
Predictors that are not numeric are ignored in the calculations.
method = "zv" identifies numeric predictor columns with a single value (i.e. having zero variance) and excludes them from further calculations. Similarly,
method = "nzv" does the same by applying
nearZeroVar with the default parameters to exclude "near zero-variance" predictors.
method = "conditionalX" examines the distribution of each predictor conditional on the outcome. If there is only one unique value within any class, the predictor is excluded from further calculations (see
checkConditionalX for an example). When
outcome is not a factor, this calculation is not executed. This operation can be time consuming when used within resampling via
The operations are applied in this order: zero-variance filter, near-zero variance filter, Box-Cox/Yeo-Johnson/exponential transformation, centering, scaling, range, imputation, PCA, ICA then spatial sign. This is a departure from versions of caret prior to version 4.76 (where imputation was done first) and is not backwards compatible if bagging was used for imputation.
If PCA is requested but centering and scaling are not, the values will still be centered and scaled. Similarly, when ICA is requested, the data are automatically centered and scaled.
k-nearest neighbor imputation is carried out by finding the k closest samples (Euclidian distance) in the training set. Imputation via bagging fits a bagged tree model for each predictor (as a function of all the others). This method is simple, accurate and accepts missing values, but it has much higher computational cost. Imputation via medians takes the median of each predictor in the training set, and uses them to fill missing values. This method is simple, fast, and accepts missing values, but treats each predictor independently, and may be inaccurate.
A warning is thrown if both PCA and ICA are requested. ICA, as implemented by the
fastICA package automatically does a PCA decomposition prior to finding the ICA scores.
The function will throw an error of any numeric variables in
x has less than two unique values unless either
method = "zv" or
method = "nzv" are invoked.
Non-numeric data will not be pre-processed and there values will be in the data frame produced by the
predict function. Note that when PCA or ICA is used, the non-numeric columns may be in different positions when predicted.
- the function call
- a named list of operations and the variables used for each
- the dimensions of
- Box-Cox transformation values, see
- a vector of means (if centering was requested)
- a vector of standard deviations (if scaling or PCA was requested)
- a matrix of eigenvectors if PCA was requested
- the value of
- the value of
- a matrix of min and max values for each predictor when
methodincludes "range" (and
- the number of principal components required of capture the specified amount of variance
- contains values for the
Kmatrix of the decomposition
- a vector of medians (if median imputation was requested)
preProcessresults in a list with elements
predict.preProcesswill produce a data frame.
Kuhn and Johnson (2013), Applied Predictive Modeling, Springer, New York (chapter 4)
Kuhn (2008), Building predictive models in R using the caret (http://www.jstatsoft.org/article/view/v028i05/v28i05.pdf)
Box, G. E. P. and Cox, D. R. (1964) An analysis of transformations (with discussion). Journal of the Royal Statistical Society B, 26, 211-252.
Box, G. E. P. and Tidwell, P. W. (1962) Transformation of the independent variables. Technometrics 4, 531-550.
Manly, B. L. (1976) Exponential data transformations. The Statistician, 25, 37 - 42.
Yeo, I-K. and Johnson, R. (2000). A new family of power transformations to improve normality or symmetry. Biometrika, 87, 954-959.
data(BloodBrain) # one variable has one unique value ## Not run: # preProc <- preProcess(bbbDescr) # # preProc <- preProcess(bbbDescr[1:100,-3]) # training <- predict(preProc, bbbDescr[1:100,-3]) # test <- predict(preProc, bbbDescr[101:208,-3]) # ## End(Not run)