privateEC: Private Evaporative Cooling feature selection and...

In insilico/privateEC: Privacy Preserving Evaporative Cooling Feature Selection and Classification with Relief-F and Random Forests

Description

Private Evaporative Cooling feature selection and classification

Usage

privateEC(
  train.ds = NULL,
  holdout.ds = NULL,
  validation.ds = NULL,
  label = "class",
  method.model = "classification",
  is.simulated = TRUE,
  bias = 0.4,
  update.freq = 5,
  importance.name = "relieff",
  importance.algorithm = "ReliefFequalK",
  relief.k.method = "k_half_sigma",
  learner.name = "randomforest",
  xgb.obj = "binary:logistic",
  use.nestedCV = FALSE,
  ncv_folds = c(10, 10),
  learner.cv = NULL,
  rf.mtry = NULL,
  rf.ntree = 500,
  xgb.num.rounds = c(1),
  xgb.max.depth = c(4),
  xgb.shrinkage = c(1),
  start.temp = 0.1,
  final.temp = 1e-05,
  tau.param = 100,
  threshold = 4/sqrt(nrow(train.ds)),
  tolerance = 1/sqrt(nrow(train.ds)),
  signal.names = NULL,
  save.file = NULL,
  verbose = FALSE
)

Arguments

train.ds	A data frame with training data and outcome labels
holdout.ds	A data frame with holdout data and outcome labels
validation.ds	A data frame with validation data and outcome labels
label	A character vector of the outcome variable column name.
method.model	Column name of outcome variable (string), classification or regression. If the analysis goal is classification make the column a factor type. For regression, make outcome column numeric type.
is.simulated	Is the data simulated (or real?)
bias	A numeric for effect size in simulated signal variables
update.freq	An integer the number of steps before update
importance.name	A character vector containg the importance algorithm name
importance.algorithm	A character vestor containing a specific importance algorithm subtype
relief.k.method	A character of numeric to indicate number of nearest neighbors for relief algorithm. Possible characters are: k_half_sigma (floor((num.samp-1)*0.154)), m6 (floor(num.samp/6)), myopic (floor((num.samp-1)/2)), and m4 (floor(num.samp/4))
learner.name	A character vector containg the learner algorithm name
xgb.obj	A character vector containing the XGBoost ojective function name
use.nestedCV	A logic character indicating whether use nested cross validation or not
ncv_folds	A vector of integers fo the number of nested cross validation folds
learner.cv	An integer for the number of cross validation folds
rf.mtry	An integer for the number of variables used for node splits
rf.ntree	An integer the number of trees in the random forest
xgb.num.rounds	= A vector of integers for xgboost algorithm iterations
xgb.max.depth	A vector of integers for the xboost maximum tree depth
xgb.shrinkage	= A vector of numerics for xgboost shrinkage values 0-1
start.temp	A numeric EC starting temperature
final.temp	A numeric EC final temperature
tau.param	A numeric tau to control temperature reduction schedule
threshold	A numeric, default 4 / sqrt(n) suggested in the thresholdout’s supplementary material (Dwork, et al.,2015)
tolerance	A numeric, default 1 / sqrt(n) suggested in the thresholdout’s supplementary material (Dwork, et al.,2015)
signal.names	A character vector of signal names in simulated data
save.file	A character vector for results filename or NULL to skip
verbose	A flag indicating whether verbose output be sent to stdout

Value

A list with:

algo.acc

data frame of results, a row for each update

ggplot.data

melted results data frame for plotting with ggplot

correct

number of variables detected correctly in each data set

atts.remain

name of the attributes in each iteraction

ncv.atts

name of the selected attributes using nested cross validation

elapsed

total elapsed time

Note

Within thresholdout, we choose a threshold of 4 / sqrt(n) and tolerance of 1 / sqrt(n) as suggested in the thresholdout’s supplementary material (Dwork, et al., 2015).

References

Trang Le, W. K. Simmons, M. Misaki, B.C. White, J. Savitz, J. Bodurka, and B. A. McKinney. “Differential privacy-based Evaporative Cooling feature selection and classification with Relief-F and Random Forests,” Bioinformatics. Accepted. https://doi.org/10.1093/bioinformatics/btx298. 2017

For more information see: Insilico Lab privateEC Page

See Also

Other classification: epistasisRank(), getImportanceScores(), originalThresholdout(), privateRF(), standardRF(), xgboostRF()

Examples

num.samples <- 100
num.variables <- 100
pct.signals <- 0.1
label <- "class"
sim.data <- createSimulation(num.samples = num.samples,
                             num.variables = num.variables,
                             pct.signals = pct.signals,
                             label = label,
                             pct.train = 1 / 3,
                             pct.holdout = 1 / 3,
                             pct.validation = 1 /3,
                             sim.type = "mainEffect",
                             verbose = FALSE)
pec.results <- privateEC(train.ds = sim.data$train,
                         holdout.ds = sim.data$holdout,
                         validation.ds = sim.data$validation,
                         label = sim.data$label,
                         is.simulated = TRUE,
                         importance.name = "relieff",
                         learner.name = "randomforest",
                         signal.names = sim.data$signal.names,
                         verbose = FALSE)
pec.results <- privateEC(train.ds = sim.data$train,
                         holdout.ds = sim.data$holdout,
                         validation.ds = sim.data$validation,
                         label = sim.data$label,
                         is.simulated = TRUE,
                         learner.name = "xgboost",
                         xgb.max.depth = 5,
                         signal.names = sim.data$signal.names,
                         verbose = FALSE)

insilico/privateEC documentation built on May 22, 2020, 5:12 p.m.