predict.rfArb: predict method for rfArb result

In Rborist: Extensible, Parallelizable Implementation of the Random Forest Algorithm

predict method for rfArb result

Description

Prediction and test using Rborist.

Usage

## S3 method for class 'rfArb'
predict(object, newdata, sampler, yTest=NULL,
keyedFrame = FALSE, quantVec=numeric(0), quantiles = length(quantVec) > 0,
ctgCensus = "votes", indexing = FALSE, trapUnobserved = FALSE,
bagging = FALSE, nThread = 0, verbose = FALSE, ...)

Arguments

object	an object of class rfArb, created from a previous invocation of the command rfArb or Rborist to train.
newdata	a design frame or matrix containing new data, with the same signature of predictors as in the training command.
sampler	an object of class Sampler used in the command.
yTest	a response vector against which to test the new predictions.
keyedFrame	whether the columns of newdata may appear in arbitrary order or as a superset of the predictors used to train.
quantVec	a vector of quantiles to predict.
quantiles	whether to predict quantiles.
ctgCensus	whether/how to summarize per-category predictions. "votes" specifies the number of trees predicting a given class. "prob" specifies a normalized, probabilistic summary. "probSample" specifies sample-weighted probabilities, similar to quantile histogramming.
indexing	whether to record the final node index, typically terminal, of tree traversal.
trapUnobserved	reports score for nonterminal upon encountering values not observed during training, such as missing data.
bagging	whether prediction is restricted to out-of-bag samples.
nThread	suggests ans OpenMP-style thread count. Zero denotes default processor setting.
verbose	whether to output progress of prediction.
...	not currently used.

Value

an object of one of two classes:

• SummaryReg summarizing regression, consisting of:

  • prediction an object of class PredictReg consisting of:

    • yPred the estimated numerical response.

    • qPred quantiles of prediction, if requested.

    • qEst quantile of the estimate, if quantiles requested.

    • indices final index of prediction, if requested.

  • validation if validation requested, an object of class ValidReg consisting of:

    • mse the mean-squared error of the estimate.

    • rsq the r-squared statistic of the estimate.

    • mae the mean absolute error of the estimate.

  • importance if permution importance requested, an object of class importanceReg, containing multiple instances of:

    • names the predictor names.

    • mse the per-predictor mean-squared error, under permutation.

• SummaryCtg summarizing classification, consisting of:

  • PredictCtg consisting of:

    • yPred estimated categorical response.

    • census factor-valued matrix of the estimate, by category, if requested.

    • prob matrix of estimate probabilities, by category, if requested.

    • indices final index of prediction, if requested.

  • validation if validation requested, an object of class ValidCtg consisting of:

    • confusion the confusion matrix.

    • misprediction the misprediction rate.

    • oobError the out-of-bag error.

  • importance if permution importance requested, an object of class importanceCtg, consisting of:

    • mispred the misprediction rate, by predictor.

    • oobErr the out-of-bag error, by predictor.

Author(s)

Mark Seligman at Suiji.

See Also

rfTrain

Examples

## Not run: 
  # Regression example:
  nRow <- 5000
  x <- data.frame(replicate(6, rnorm(nRow)))
  y <- with(x, X1^2 + sin(X2) + X3 * X4) # courtesy of S. Welling.

  pf <- preformat(x)
  sp <- presample(y)
  rb <- rfArb(pf, sp, y)


  # Performs separate prediction on new data:
  xx <- data.frame(replace(6, rnorm(nRow)))
  pred <- predict(rb, xx)
  yPred <- pred$yPred

  rb <- Rborist(x,y)

  # Performs separate prediction on new data:
  xx <- data.frame(replacate(6, rnorm(nRow)))
  pred <- predict(rb, xx)
  yPred <- pred$yPred

  # As above, but also records final indices of each tree walk:
  #
  pred <- predict(rb, xx, indexing=TRUE)
  print(pred$indices[c(1:2), ])


  # As above, but predicts over \code{newdata} with unobserved values.
  # In the case of numerical data, only missing values are considered
  # unobserved.  Missing values are encoded as \code{NaN}, which are
  # incomparable, precipitating \code{false} on every test.  Prediction
  # therefore takes the \code{false} branch when encountering missing
  # values:
  #
  xxMissing <- xx
  xxMissing[6, c(15, 32, 87, 101)] <- NA
  pred <- predict(rb, xxMissing)
  

  # As above, but returns a nonterminal score upon encountering
  # unobserved values. Neither the true nor the false branch from the
  # testing node is taken.  Instead, the score returned is derived
  # from all leaf nodes (terminals) reached by the testing
  # (nonterminal) node.
  #
  pred <- predict(rb, xxMissing, trapUnobserved = TRUE)


  # Performs separate prediction, using original response as test
  # vector:
  pred <- predict(rb, xx, y)
  mse <- pred$mse
  rsq <- pred$rsq


  # Performs separate prediction with (default) quantiles:
  pred <- predict(rb, xx, quantiles="TRUE")
  qPred <- pred$qPred


  # Performs separate prediction with deciles:
  pred <- predict(rb, xx, quantVec = seq(0.1, 1.0, by = 0.10))
  qPred <- pred$qPred


  # Classification examples:
  data(iris)
  rb <- Rborist(iris[-5], iris[5])


  # Generic prediction using training set.
  # Census as (default) votes:
  pred <- predict(rb, iris[-5])
  yPred <- pred$yPred
  census <- pred$census

  # Using the \code{keyedFrame} option allows the columns of
  # \code{newdata} to appear in arbitrary order, so long as the
  # columns present during training appear as a subset:
  #
  pred <- predict(rb, iris[c(2, 4, 3, 1)], keyedFrame=TRUE)


  # As above, but validation census to report class probabilities:
  pred <- predict(rb, iris[-5], ctgCensus="prob")
  prob <- pred$prob


  # As above, but with training reponse as test vector:
  pred <- predict(rb, iris[-5], iris[5], ctgCensus = "prob")
  prob <- pred$prob
  conf <- pred$confusion
  misPred <- pred$misPred

  # As above, but predicts nonterminal when encountering categories
  # not observed during training.  That is, prediction returns a score
  # derived from all terminal nodes (leaves) reached from the
  # (nonterminal) testing node.
  #
  # In this case, "unobserved" refers to categories not present in
  # the subpartition over which a splitting is performed.  As training
  # partitions the data into smaller and smaller regions, a given
  # category becomes less likely to appear in a region.
  #
  # More generally, unobserved data can include missing predictors as
  # well as categories appearing in \code{newdata} which were not
  # present during training.
  #
  pred <- predict(rb, trapUnobserved=TRUE)

## End(Not run)

Rborist documentation built on April 3, 2025, 8:04 p.m.