AUCRF (version 1.1)

AUCRF: Variable Selection with Random Forest and the Area Under the Curve

Description

AUCRF is an algorithm for variable selection using Random Forest based on optimizing the area-under-the ROC curve (AUC) of the Random Forest. The proposed strategy implements a backward elimination process based on the initial ranking of the variables.

Usage

AUCRF(formula, data, k0 = 1, pdel = 0.2, ranking=c("MDG","MDA"), ...)

Arguments

formula

an object of class formula: a symbolic description of the model to be fitted. The details of model specification are given in Details.

data

a data frame containing the variables in the model. Dependent variable must be a binary variable defined as factor and codified as 1 for positives (e.g. cases) and 0 for negatives (e.g. controls).

k0

number of remaining variables for stopping the backward elimination process. By default k0=1.

pdel

fraction of remaining variables to be removed in each step. By default pdel=0.2. If pdel=0, only one variable is removed each time.

ranking

specifies the importance measure provided by randomForest for ranking the variables. There are two options MDG (by default) for MeanDecreaseGini and MDA for MeanDecreaseAccuracy.

...

optional parameters to be passed to the randomForest function. If no arguments are specified, default arguments of randomForest function will be used.

Value

An object of class AUCRF, which is a list with the following components:

call

the original call to AUCRF.

data

the data argument.

ranking

the ranking of predictors based on the importance measure.

Xopt

optimal set of predictors obtained.

OOB-AUCopt

AUC obtained for the optimal set of predictors.

Kopt

size of the optimal set of predictors obtained.

AUCcurve

values of AUC obtained for each set of predictors evaluated in the elimination process.

RFopt

the randomForest adjusted with the optimal set.

Details

The AUC-RF algorithm is described in detail in Calle et. al.(2011). The following is a summary: Ranking and AUC of the initial set: Perform a random forest using all predictor variables and the response, as specified in the formula argument, and compute the AUC of the random forest. Based on the selected measure of importance (by default MDG), obtain a ranking of predictors. Elimination process: Based on the variables ranking, remove the less important variables (fraction of variables specified in pdel argument). Perform a new random forest with the remaining variables and compute its AUC. This step is iterated until the number of remaining variables is less or equal than k0. Optimal set: The optimal set of predictive variables is considered the one giving rise to the Random Forest with the highest OOB-AUC$opt$. The number of selected predictors is denoted by K$opt$

References

Calle ML, Urrea V, Boulesteix A-L, Malats N (2011) "AUC-RF: A new strategy for genomic profiling with Random Forest". Human Heredity. (In press)

See Also

OptimalSet, AUCRFcv, randomForest.

Examples

   
  # load the included example dataset. This is a simulated case/control study  
  # data set with 4000 patients (2000 cases / 2000 controls) and 1000 SNPs, 
  # where the  first 10 SNPs have a direct association with the outcome:
  data(exampleData)
  
  # call AUCRF process: (it may take some time)
  # fit <- AUCRF(Y~., data=exampleData)
  
  # The result of this example is included for illustration purpose:
  
  data(fit)
  summary(fit)
  plot(fit)
  
  # Additional randomForest parameters can be included, otherwise default
  # parameters of randomForest function will be used:
  # fit <- AUCRF(Y~., data=exampleData, ntree=1000, nodesize=20)