Partager via

rxPredict: Predicted values and residuals for model objects built using RevoScaleR

  • Article

Description

Compute predicted values and residuals using the following objects: rxLinMod, rxGlm, rxLogit, rxDTree, rxBTrees, and rxDForest.

Usage

  rxPredict(modelObject, data = NULL, ...)
 ## S3 method for class `default':
rxPredict  (modelObject, data = NULL, outData = NULL,     
            computeStdErrors = FALSE, interval = "none", confLevel = 0.95,
            computeResiduals = FALSE, type = c("response", "link"), 
            writeModelVars = FALSE, extraVarsToWrite = NULL, removeMissings = FALSE,
            append = c("none", "rows"), overwrite = FALSE, checkFactorLevels = TRUE,
            predVarNames = NULL, residVarNames = NULL, 
            intervalVarNames = NULL, stdErrorsVarNames = NULL, predNames = NULL, 
            blocksPerRead = rxGetOption("blocksPerRead"),
            reportProgress = rxGetOption("reportProgress"), verbose = 0,
            xdfCompressionLevel = rxGetOption("xdfCompressionLevel"), ...)

Arguments

modelObject

object returned from a RevoScaleR model fitting function. Valid values include rxLinMod, rxLogit, rxGlm, rxDTree, rxBTrees, and rxDForest. Objects with multiple dependent variables are not supported in rxPredict.

data

An RxXdfData data source object to be used for predictions. If not using a distributed compute context such as RxHadoopMR, a data frame, or a character string specifying the input .xdf file can also be used.

outData

file or existing data frame to store predictions; can be same as the input file or NULL. If not NULL, a character string specifying the output �.xdf� file, a RxXdfData object, a RxOdbcData data source, or a RxSqlServerData data source. outData can also be a delimited RxTextData data source if using a native file system and not appending.

computeStdErrors

logical value. If TRUE, the standard errors for each dependent variable are calculated.

interval

character string defining the type of interval calculation to perform. Supported values are "none", "confidence", and "prediction".

confLevel

numeric value representing the confidence level on the interval [0, 1].

computeResiduals

logical value. If TRUE, residuals are computed.

type

Applies to rxGlm and rxLogit, used to set the type of prediction. Valid values are "response"and "link". If type = "response", the predictions are on the scale of the response variable. For instance, for the binomial model, the predictions are in the range (0,1). If type = "link", the predictions are on the scale of the linear predictors. Thus for the binomial model, the predictions are of log-odds.

writeModelVars

logical value. If TRUE, and the output data set is different from the input data set, variables in the model will be written to the output data set in addition to the predictions (and residuals, standard errors, and confidence bounds, if requested). If variables from the input data set are transformed in the model, the transformed variables will also be included.

extraVarsToWrite

NULL or character vector of additional variables names from the input data or transforms to include in the outData. If writeModelVars is TRUE, model variables will be included as well.

removeMissings

logical value. If TRUE, rows with missing values are removed.

append

either "none" to create a new files or "rows" to append rows to an existing file. If outData exists and append is "none", the overwrite argument must be set to TRUE. You can append only to RxTeradata data source. Ignored for data frames.

overwrite

logical value. If TRUE, an existing outData will be overwritten. overwrite is ignored if appending rows. Ignored for data frames.

checkFactorLevels

logical value. If TRUE, up to 1000 factor levels for the data will be verified against factor levels in the model. Setting to FALSE can speed up computations if using lots of factors.

predVarNames

character vector specifying name(s) to give to the prediction results

residVarNames

character vector specifying name(s) to give to the residual results.

intervalVarNames

NULL or a character vector defining low and high confidence interval variable names, respectively. If NULL, the strings "_Lower" and "_Upper" are appended to the dependent variable names to form the confidence interval variable names.

stdErrorsVarNames

NULL or a character vector defining variable names corresponding to the standard errors, if calculated. If NULL, the string "_StdErr" is appended to the dependent variable names to form the standard errors variable names.

predNames

character vector specifying name(s) to give to the prediction and residual results; if length is 2, the second name is used for residuals. This argument is deprecated and predVarNames and residVarNames should be used instead.

blocksPerRead

number of blocks to read for each chunk of data read from the data source. If the data and outData are the same file, blocksPerRead must be 1.

reportProgress

integer value with options:

  • 0: no progress is reported.
  • 1: the number of processed rows is printed and updated.
  • 2: rows processed and timings are reported.
  • 3: rows processed and all timings are reported.

verbose

integer value. If 0, no additional output is printed. If 1, additional summary information is printed.

xdfCompressionLevel

integer in the range of -1 to 9 indicating the compression level for the output data if written to an .xdf file. The higher the value, the greater the amount of compression - resulting in smaller files but a longer time to create them. If xdfCompressionLevel is set to 0, there will be no compression and files will be compatible with the 6.0 release of Revolution R Enterprise. If set to -1, a default level of compression will be used.

...

additional arguments to be passed directly to the Revolution Compute Engine.

Details

rxPredict computes predicted values and/or residuals from an existing model type. The most common way to call rxPredict is rxPredict(modelObject, data, outData). Typically, all the other arguments are left at their defaults.

For rxLogit, the residuals are equivalent to those computed for glm with type set to "response", e.g., residuals(glmObj, type="response").

If the data is the same data used to create the modelObject, the predicted values are the fitted values for the original model.

If the data specified is an .xdf file, the outData must be NULL or an .xdf file. If outData is an .xdf file, the computed data will be appended as columns. If outData is NULL, the computed columns will be appended to the data .xdf file.

If the data specified is a data frame, the outData must be NULL or a data frame. If outData is a data frame, a copy of the data frame with the new columns appended will be returned. If outData is NULL, a vector or list of the computed values will be returned.

If a transformation function is being used for the model estimation, the information variable .rxIsPrediction can be used to exclude computations for the dependent variable when running rxPredict. See rxTransform for an example.

Value

If a data frame is specified as the input data, a data frame is returned. If a data frame is specified as the outData, variables containing the results are added to the data frame and it is returned. If outData is NULL, a data frame containing the predicted values (and residuals and standard errors, if requested) is returned.

If an .xdf file is specified as the input data, an RxXdfData data source object is returned that can be used in subsequent RevoScaleR analyses. If outData is an .xdf file, the RxXdfData data source represents the outData file. If outData is NULL, the predicted values (and, if requested, residuals) are appended to the original data file. The returned RxXdfData object represents this file.

Computing Standard Errors of Predicted Values

Use computeStdErrors to control whether or not prediction standard errors are computed.

Use interval to control whether confidence or prediction (tolerance) intervals are computed at the specified level (confLevel). These are sometimes referred to as narrow and wide intervals, respectively.

Use stdErrorsVarNames to name the standard errors output variable and intervalVarNames to specify the output variable names of the lower and upper confidence/tolerance intervals.

In calculating the prediction standard errors, keep the following in mind:

* Prediction standard errors are available for both rxLinMod and rxLogit models.

* Standard errors are computationally intensive for large models, i.e., those involving a large number of model parameters.

* rxLinMod and rxLogit must be called with covCoef = TRUE because the variance-covariance matrix of the coefficients must be available.

* Cube regressions are not supported (cube = TRUE).

* Multiple dependent variables are currently not supported.

* For rxLogit, interval = "confidence" is supported (unlike predict.glm, which does not support confidence bounds), but interval = "prediction" is not supported.

* If residuals are requested, and if there are missing values in the dependent variable, then all computed values (prediction, standard errors, confidence levels) will be assigned the value missing, and will be removed if removeMissings = TRUE. If no residuals are requested, then missings in the dependent variable (which need not exist in the data) have no effect.

Author(s)

Microsoft Corporation Microsoft Technical Support

See Also

rxLinMod, rxLogit, rxGlm, rxPredict.rxDTree, rxPredict.rxDForest, rxPredict.rxNaiveBayes.

Examples


 # Load the built-in iris data set and predict sepal length
 myIris <- iris
 myIris[1:5,]
 form <- Sepal.Length ~ Sepal.Width + Petal.Length + Petal.Width + Species
 irisLinMod <- rxLinMod(form, data = myIris)
 myIrisPred <- rxPredict(modelObject = irisLinMod, data = myIris)
 myIris$SepalLengthPred <- myIrisPred$Sepal.Length_Pred
 myIris[1:5,]
 irisResiduals<- rxPredict(modelObject = irisLinMod, data = myIris, computeResiduals = TRUE)
 names(irisResiduals)

 # Use sample data to compare lm and glm results with rxPredict
 sampleDataDir <- rxGetOption("sampleDataDir")
 mortFile <- file.path(sampleDataDir, "mortDefaultSmall.xdf")
 linModPredictFile <- file.path(tempdir(), "mortPredictLinMod.xdf")
 logitPredictFile <- file.path(tempdir(), "mortPredictLogit.xdf")
 mortDF <- rxDataStep(inData = mortFile)

 # Compare residuals from rxLinMod with lm
 linMod <- rxLinMod(creditScore ~ yearsEmploy, data = mortFile)
 rxPredict(modelObject = linMod, data = mortFile, outData = linModPredictFile,
           computeResiduals = TRUE)
 residDF <- rxDataStep(inData = linModPredictFile)
 mortLM <- lm(creditScore ~ yearsEmploy, data = mortDF)
 # Sum of differences should be very small
 sum(mortLM$residuals - residDF$creditScore_Resid)

 # Create logit model object and compute predictions and residuals
 logitModObj <- rxLogit(default ~ creditScore, data = mortFile)
 rxPredict(modelObject = logitModObj, data = mortFile,
           outData = logitPredictFile, computeResiduals = TRUE)
 residDF <- rxDataStep(inData = logitPredictFile)
 mortGLM <- glm(default ~ creditScore, data = mortDF, family = binomial())

 # maximum differences should be very small
 max(abs(mortGLM$fitted.values - residDF$default_Pred)) 
 max(abs(residuals(mortGLM, type = "response") - residDF$default_Resid))