Post-processing for Regression
Post-processing use OOB votes and predicted values to build more accurate estimates of Response values. Note that post-processing can not ensure that new estimate will have a lower error. It works for many cases but not all.
postProcessingVotes(object, nbModels = 1, idx = 1, granularity = 1, predObject = NULL, swapPredictions = FALSE, X = NULL, Xtest = NULL, imbalanced = FALSE, OOB = FALSE, method = c("cutoff", "bias", "residuals"), keep2ndModel = FALSE, largeData = FALSE, ...)
object: a randomUniformForest object with OOB data.nbModels: how many models to build for new estimates. Usually one is enough.idx: how many values to choose in OOB model for each new predicted value. Usually one is enough.granularity: degree of precision needed for each old estimate value. Usually one is enough.predObject: if current model is built with full sample, then using an old model 'predObject' (a randomUniformForest object) that have OOB data can help to reduce error. Must be used with 'swapPredictions = TRUE'swapPredictions: set it to TRUE if two models, current one without OOB data and old one with OOB data, have to be used for trying to reduce prediction error.X: not currently used.Xtest: test data in the case of regression, for a more friendly output of the model.imbalanced: if TRUE, may improve metrics in the case of imbalanced datasets.OOB: if FALSE, does not use OOB informations.method: for classification, if expected bias is enough high, one may use it as a method to improve AUC. Otherwise, use the default one, 'cutoff'. Both tend to get the same results, despite a few tests for now, but 'bias' method seems more robust. 'residuals' is used in regression only as a powerful but computationally intensive method and replaces the default internal one.keep2ndModel: if TRUE, and for regression, keep the model based on residuals for further modelling and predictions.largeData: if TRUE, and for regression, use rUniformForest.big to compute model for the residuals....: arguments to use for the computation of the model from the residuals.Xu, Ruo, Improvements to random forest methodology (2013). Graduate Theses and Dissertations. Paper 13052.
a vector of predicted values.
Saip Ciss saip.ciss@wanadoo.fr
# Note that post-processing works better with enough trees, at least 100, and enough data n = 200; p = 20 # Simulate 'p' gaussian vectors with random parameters between -10 and 10. X <- simulationData(n,p) # give names to features X <- fillVariablesNames(X) # Make a rule to create response vector epsilon1 = runif(n,-1,1) epsilon2 = runif(n,-1,1) # a rule with many noise (only four variables are significant) Y = 2*(X[,1]*X[,2] + X[,3]*X[,4]) + epsilon1*X[,5] + epsilon2*X[,6] # randomize then make train and test sample twoSamples <- cut(sample(1:n,n), 2, labels = FALSE) Xtrain = X[which(twoSamples == 1),] Ytrain = Y[which(twoSamples == 1)] Xtest = X[which(twoSamples == 2),] Ytest = Y[which(twoSamples == 2)] # compute an accurate model (in this case bagging and log data works best) and predict rUF.model <- randomUniformForest(Xtrain, Ytrain, xtest = Xtest, ytest = Ytest, bagging = TRUE, logX = TRUE, ntree = 60, threads = 2) # get mean squared error rUF.model # post process newEstimate <- postProcessingVotes(rUF.model) # get mean squared error sum((newEstimate - Ytest)^2)/length(Ytest) ## regression do not use all data but sub-samples. ## Comparing, when using full sample (but then, we do not have OOB data) # rUF.model.fullsample <- randomUniformForest(Xtrain, Ytrain, xtest = Xtest, ytest = Ytest, # subsamplerate = 1, bagging = TRUE, logX = TRUE, ntree = 60, threads = 2) # rUF.model.fullsample ## Nevertheless we can use old model with OOB data to fit a new estimate ## newEstimate.fullsample <- postProcessingVotes(rUF.model.fullsample, # predObject = rUF.model, swapPredictions = TRUE) ## get mean squared error # sum((newEstimate.fullsample - Ytest)^2)/length(Ytest)
Useful links