Define a learner — learner • timesift

A learner is a pair of functions: one that fits a model to a representation and a response, and one that predicts from it. Everything the package fits goes through this pair, so a learner of your own sits beside the ones that ship and needs no change to the ladder, the folds or the scoring.

Usage

learner(
  name,
  fit,
  predict,
  data = NULL,
  reads = c("tabular", "sequence"),
  multi = c("separate", "joint"),
  control = NULL,
  needs = character(),
  params = list()
)

Arguments

name: Name the learner is reported under.
fit: A function of (x, y, ...), where x is a [unit, bin, channel] array and y the response matrix for the same units, returning a fitted object. A fit that declares a control argument is handed the resolved train_control(), and one that declares a head argument is handed the registered response head, whose loss and activation say what it is fitting toward. The learners that ship read both from there and hold no response of their own. One that declares a group argument is handed the grouping the outer fold map keeps whole, one value per unit of x or NULL, so a split it draws inside the fit keeps the same groups whole.
predict: A function of (model, x) returning a [unit, variable] matrix of predictions for the units of x, in that order.
data: A representation the learner is pinned to, or NULL to run across every representation of the run.
reads: "tabular" or "sequence".
multi: "separate" where the learner fits one model per response, "joint" where one model covers them all. It is handed the whole response matrix either way; this is what the learner says it does with it, and what a report says of the candidate.
control: A train_control() for this learner alone. The settings it names override the control the run was given; everything else is taken from that one.
needs: Packages the learner requires. A learner that cannot run says so at once rather than falling back to something else.
params: Settings carried with the learner and passed to fit.

Value

A timesift_learner.

Details

A learner declares what it can be handed and how it covers several responses. reads is "tabular" where the bins reach it as a block of predictors and "sequence" where their order in time is what it reads, and multi is "joint" where one fitted model covers every response and "separate" where the learner fits one per response. Either way it is handed the whole response matrix and returns one column per response, so nothing above the learner layer has to know which it was, and the block of predictors is built once for the fit rather than once for every response of it.

data pins a learner to one representation. Left NULL the learner runs across every representation of the run.

Examples

# The bin means of a unit, fed to one logistic regression per variable.
flat_glm <- learner(
  "flat_glm",
  fit = function(x, y, ...) {
    f <- as.data.frame(apply(x, c(1, 3), mean))
    lapply(seq_len(ncol(y)), function(j)
      stats::glm(y[, j] ~ ., data = f, family = stats::binomial()))
  },
  predict = function(model, x) {
    f <- as.data.frame(apply(x, c(1, 3), mean))
    vapply(model, function(m) stats::predict(m, f, type = "response"), numeric(nrow(f)))
  }
)
flat_glm