R: Fit an Artificial Neural Network model on "tabular" provided...

ann_tab_cv {glmnetr}

R Documentation

Fit an Artificial Neural Network model on "tabular" provided as a matrix, optionally allowing for an offset term

Description

Fit an Artificial Neural Network model for analysis of "tabular" data. The model has two hidden layers where the number of terms in each layer is configurable by the user. The activation function can also be switched between relu() (default) gelu() or sigmoid(). Optionally an offset term may be included. Model "family" may be "cox" to fit a generalization of the Cox proportional hazards model, "binomial" to fit a generalization of the logistic regression model and "gaussian" to fit a generalization of linear regression model for a quantitative response. See the corresponding vignette for examples.

Usage

ann_tab_cv(
  myxs,
  mystart = NULL,
  myy,
  myevent = NULL,
  myoffset = NULL,
  family = "binomial",
  fold_n = 5,
  epochs = 200,
  eppr = 40,
  lenz1 = 16,
  lenz2 = 8,
  actv = 1,
  drpot = 0,
  mylr = 0.005,
  wd = 0,
  l1 = 0,
  lasso = 0,
  lscale = 5,
  scale = 1,
  resetlw = 1,
  minloss = 1,
  gotoend = 0,
  seed = NULL,
  foldid = NULL
)

Arguments

`myxs`	predictor input - an n by p matrix, where n (rows) is sample size, and p (columns) the number of predictors. Must be in matrix form for complete data, no NA's, no Inf's, etc., and not a data frame.
`mystart`	an optional vector of start times in case of a Cox model. Class numeric of length same as number of patients (n)
`myy`	dependent variable as a vector: time, or stop time for Cox model, Y_ 0 or 1 for binomial (logistic), numeric for gaussian. Must be a vector of length same as number of sample size.
`myevent`	event indicator, 1 for event, 0 for census, Cox model only. Must be a numeric vector of length same as sample size.
`myoffset`	an offset term to be used when fitting the ANN. Not yet implemented in its pure form. Functionally an offset can be included in the first column of the predictor or feature matrix myxs and indicated as such using the lasso option.
`family`	model family, "cox", "binomial" or "gaussian" (default)
`fold_n`	number of folds for each level of cross validation
`epochs`	number of epochs to run when tuning on number of epochs for fitting final model number of epochs informed by cross validation
`eppr`	for EPoch PRint. print summary info every eppr epochs. 0 will print first and last epochs, 0 for first and last epoch, -1 for minimal and -2 for none.
`lenz1`	length of the first hidden layer in the neural network, default 16
`lenz2`	length of the second hidden layer in the neural network, default 16
`actv`	for ACTiVation function. Activation function between layers, 1 for relu, 2 for gelu, 3 for sigmoid.
`drpot`	fraction of weights to randomly zero out. NOT YET implemented.
`mylr`	learning rate for the optimization step in the neural network model fit
`wd`	a possible weight decay for the model fit, default 0 for not considered
`l1`	a possible L1 penalty weight for the model fit, default 0 for not considered
`lasso`	1 to indicate the first column of the input matrix is an offset term, often derived from a lasso model, else 0 (default)
`lscale`	Scale used to allow ReLU to exend +/- lscale before capping the inputted linear estimated
`scale`	Scale used to transform the inital random paramter assingments by dividing by scale
`resetlw`	1 as default to re-adjust weights to account for the offset every epoch. This is only used in case lasso is set to 1.
`minloss`	default of 1 for minimizing loss, else maximizing agreement (concordance for Cox and Binomial, R-square for Gaussian), as function of epochs by cross validaition
`gotoend`	fit to the end of epochs. Good for plotting and exploration
`seed`	an optional a numerical/integer vector of length 2, for R and torch random generators, default NULL to generate these. Integers should be positive and not more than 2147483647.
`foldid`	a vector of integers to associate each record to a fold. Should be integers from 1 and fold_n.

Value

an artificial neural network model fit

Author(s)