SensAnalysisMLP: Sensitivity of MLP models

In NeuralSens: Sensitivity Analysis of Neural Networks

Sensitivity of MLP models

Description

Function for evaluating the sensitivities of the inputs variables in a mlp model

Usage

SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  ...
)

## Default S3 method:
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  trData,
  actfunc = NULL,
  deractfunc = NULL,
  preProc = NULL,
  terms = NULL,
  output_name = NULL,
  ...
)

## S3 method for class 'train'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  boot.R = NULL,
  boot.seed = 1,
  boot.alpha = 0.05,
  ...
)

## S3 method for class 'H2OMultinomialModel'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  ...
)

## S3 method for class 'H2ORegressionModel'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  ...
)

## S3 method for class 'list'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  trData,
  actfunc,
  ...
)

## S3 method for class 'mlp'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  trData,
  preProc = NULL,
  terms = NULL,
  ...
)

## S3 method for class 'nn'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  preProc = NULL,
  terms = NULL,
  ...
)

## S3 method for class 'nnet'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  trData,
  preProc = NULL,
  terms = NULL,
  ...
)

## S3 method for class 'nnetar'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  ...
)

## S3 method for class 'numeric'
SensAnalysisMLP(
  MLP.fit,
  .returnSens = TRUE,
  plot = TRUE,
  .rawSens = FALSE,
  sens_origin_layer = 1,
  sens_end_layer = "last",
  sens_origin_input = TRUE,
  sens_end_input = FALSE,
  trData,
  actfunc = NULL,
  preProc = NULL,
  terms = NULL,
  ...
)

Arguments

MLP.fit	fitted neural network model
.returnSens	DEPRECATED
plot	logical whether or not to plot the analysis. By default is TRUE.
.rawSens	DEPRECATED
sens_origin_layer	numeric specifies the layer of neurons with respect to which the derivative must be calculated. The input layer is specified by 1 (default).
sens_end_layer	numeric specifies the layer of neurons of which the derivative is calculated. It may also be 'last' to specify the output layer (default).
sens_origin_input	logical specifies if the derivative must be calculated with respect to the inputs (TRUE) or output (FALSE) of the sens_origin_layer layer of the model. By default is TRUE.
sens_end_input	logical specifies if the derivative calculated is of the output (FALSE) or from the input (TRUE) of the sens_end_layer layer of the model. By default is FALSE.
...	additional arguments passed to or from other methods
trData	data.frame containing the data to evaluate the sensitivity of the model
actfunc	character vector indicating the activation function of each neurons layer.
deractfunc	character vector indicating the derivative of the activation function of each neurons layer.
preProc	preProcess structure applied to the training data. See also preProcess
terms	function applied to the training data to create factors. See also train
output_name	character name of the output variable in order to avoid changing the name of the output variable in trData to '.outcome'
boot.R	int Number of bootstrap samples to calculate. Used to detect significant inputs and significant non-linearities. Only available for train objects. Defaults to NULL.
boot.seed	int Seed of bootstrap evaluations.
boot.alpha	float Significance level of statistical test.

Details

In case of using an input of class factor and a package which need to enter the input data as matrix, the dummies must be created before training the neural network.

After that, the training data must be given to the function using the trData argument.

Value

SensMLP object with the sensitivity metrics and sensitivities of the MLP model passed to the function.

Plots

• Plot 1: colorful plot with the classification of the classes in a 2D map

• Plot 2: b/w plot with probability of the chosen class in a 2D map

• Plot 3: plot with the stats::predictions of the data provided

References

Pizarroso J, Portela J, Muñoz A (2022). NeuralSens: Sensitivity Analysis of Neural Networks. Journal of Statistical Software, 102(7), 1-36.

Examples

## Load data -------------------------------------------------------------------
data("DAILY_DEMAND_TR")
fdata <- DAILY_DEMAND_TR
## Parameters of the NNET ------------------------------------------------------
hidden_neurons <- 5
iters <- 100
decay <- 0.1

################################################################################
#########################  REGRESSION NNET #####################################
################################################################################
## Regression dataframe --------------------------------------------------------
# Scale the data
fdata.Reg.tr <- fdata[,2:ncol(fdata)]
fdata.Reg.tr[,3] <- fdata.Reg.tr[,3]/10
fdata.Reg.tr[,1] <- fdata.Reg.tr[,1]/1000

# Normalize the data for some models
preProc <- caret::preProcess(fdata.Reg.tr, method = c("center","scale"))
nntrData <- predict(preProc, fdata.Reg.tr)

#' ## TRAIN nnet NNET --------------------------------------------------------
# Create a formula to train NNET
form <- paste(names(fdata.Reg.tr)[2:ncol(fdata.Reg.tr)], collapse = " + ")
form <- formula(paste(names(fdata.Reg.tr)[1], form, sep = " ~ "))

set.seed(150)
nnetmod <- nnet::nnet(form,
                      data = nntrData,
                      linear.output = TRUE,
                      size = hidden_neurons,
                      decay = decay,
                      maxit = iters)
# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(nnetmod, trData = nntrData)

# Try SensAnalysisMLP to calculate sensitivities with respect to output of hidden neurones
NeuralSens::SensAnalysisMLP(nnetmod, trData = nntrData,
                             sens_origin_layer = 2,
                             sens_end_layer = "last",
                             sens_origin_input = FALSE,
                             sens_end_input = FALSE)
## Train caret NNET ------------------------------------------------------------
# Create trainControl
ctrl_tune <- caret::trainControl(method = "boot",
                                 savePredictions = FALSE,
                                 summaryFunction = caret::defaultSummary)
set.seed(150) #For replication
caretmod <- caret::train(form = DEM~.,
                              data = fdata.Reg.tr,
                              method = "nnet",
                              linout = TRUE,
                              tuneGrid = data.frame(size = 3,
                                                    decay = decay),
                              maxit = iters,
                              preProcess = c("center","scale"),
                              trControl = ctrl_tune,
                              metric = "RMSE")

# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(caretmod)

## Train h2o NNET --------------------------------------------------------------
# Create a cluster with 4 available cores
h2o::h2o.init(ip = "localhost",
              nthreads = 4)

# Reset the cluster
h2o::h2o.removeAll()
fdata_h2o <- h2o::as.h2o(x = fdata.Reg.tr, destination_frame = "fdata_h2o")

set.seed(150)
h2omod <-h2o:: h2o.deeplearning(x = names(fdata.Reg.tr)[2:ncol(fdata.Reg.tr)],
                                     y = names(fdata.Reg.tr)[1],
                                     distribution = "AUTO",
                                     training_frame = fdata_h2o,
                                     standardize = TRUE,
                                     activation = "Tanh",
                                     hidden = c(hidden_neurons),
                                     stopping_rounds = 0,
                                     epochs = iters,
                                     seed = 150,
                                     model_id = "nnet_h2o",
                                     adaptive_rate = FALSE,
                                     rate_decay = decay,
                                     export_weights_and_biases = TRUE)

# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(h2omod)

# Turn off the cluster
h2o::h2o.shutdown(prompt = FALSE)
rm(fdata_h2o)

## Train RSNNS NNET ------------------------------------------------------------
# Normalize data using RSNNS algorithms
trData <- as.data.frame(RSNNS::normalizeData(fdata.Reg.tr))
names(trData) <- names(fdata.Reg.tr)
set.seed(150)
RSNNSmod <-RSNNS::mlp(x = trData[,2:ncol(trData)],
                           y = trData[,1],
                           size = hidden_neurons,
                           linOut = TRUE,
                           learnFuncParams=c(decay),
                           maxit=iters)

# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(RSNNSmod, trData = trData, output_name = "DEM")

## USE DEFAULT METHOD ----------------------------------------------------------
NeuralSens::SensAnalysisMLP(caretmod$finalModel$wts,
                            trData = fdata.Reg.tr,
                            mlpstr = caretmod$finalModel$n,
                            coefnames = caretmod$coefnames,
                            actfun = c("linear","sigmoid","linear"),
                            output_name = "DEM")

################################################################################
#########################  CLASSIFICATION NNET #################################
################################################################################
## Regression dataframe --------------------------------------------------------
# Scale the data
fdata.Reg.cl <- fdata[,2:ncol(fdata)]
fdata.Reg.cl[,2:3] <- fdata.Reg.cl[,2:3]/10
fdata.Reg.cl[,1] <- fdata.Reg.cl[,1]/1000

# Normalize the data for some models
preProc <- caret::preProcess(fdata.Reg.cl, method = c("center","scale"))
nntrData <- predict(preProc, fdata.Reg.cl)

# Factorize the output
fdata.Reg.cl$DEM <- factor(round(fdata.Reg.cl$DEM, digits = 1))

# Normalize the data for some models
preProc <- caret::preProcess(fdata.Reg.cl, method = c("center","scale"))
nntrData <- predict(preProc, fdata.Reg.cl)

## Train caret NNET ------------------------------------------------------------
# Create trainControl
ctrl_tune <- caret::trainControl(method = "boot",
                                 savePredictions = FALSE,
                                 summaryFunction = caret::defaultSummary)
set.seed(150) #For replication
caretmod <- caret::train(form = DEM~.,
                                data = fdata.Reg.cl,
                                method = "nnet",
                                linout = FALSE,
                                tuneGrid = data.frame(size = hidden_neurons,
                                                      decay = decay),
                                maxit = iters,
                                preProcess = c("center","scale"),
                                trControl = ctrl_tune,
                                metric = "Accuracy")

# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(caretmod)

## Train h2o NNET --------------------------------------------------------------
# Create local cluster with 4 available cores
h2o::h2o.init(ip = "localhost",
              nthreads = 4)

# Reset the cluster
h2o::h2o.removeAll()
fdata_h2o <- h2o::as.h2o(x = fdata.Reg.cl, destination_frame = "fdata_h2o")

set.seed(150)
h2omod <- h2o::h2o.deeplearning(x = names(fdata.Reg.cl)[2:ncol(fdata.Reg.cl)],
                                       y = names(fdata.Reg.cl)[1],
                                       distribution = "AUTO",
                                       training_frame = fdata_h2o,
                                       standardize = TRUE,
                                       activation = "Tanh",
                                       hidden = c(hidden_neurons),
                                       stopping_rounds = 0,
                                       epochs = iters,
                                       seed = 150,
                                       model_id = "nnet_h2o",
                                       adaptive_rate = FALSE,
                                       rate_decay = decay,
                                       export_weights_and_biases = TRUE)

# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(h2omod)

# Apaga el cluster
h2o::h2o.shutdown(prompt = FALSE)
rm(fdata_h2o)

## TRAIN nnet NNET --------------------------------------------------------
# Create a formula to train NNET
form <- paste(names(fdata.Reg.tr)[2:ncol(fdata.Reg.tr)], collapse = " + ")
form <- formula(paste(names(fdata.Reg.tr)[1], form, sep = " ~ "))

set.seed(150)
nnetmod <- nnet::nnet(form,
                      data = nntrData,
                      linear.output = TRUE,
                      size = hidden_neurons,
                      decay = decay,
                      maxit = iters)
# Try SensAnalysisMLP
NeuralSens::SensAnalysisMLP(nnetmod, trData = nntrData)

NeuralSens documentation built on June 22, 2024, 12:06 p.m.