Constraint: predictions must be monotonic in a variable

Machine-checked, model-agnostic (ICE-style): for a sample of observed rows, the variable is swept over its observed range with everything else held fixed, and predictions must move in the stated direction every time.

Usage

con_monotone(
  var,
  direction = c("increasing", "decreasing"),
  tol = 1e-08,
  n_grid = 25,
  n_rows = 20
)

Arguments

var

Column name (numeric) the response must be monotonic in.

direction

"increasing" or "decreasing".

tol

Tolerance for tiny numeric wiggles.

n_grid, n_rows

Size of the sweep grid and number of rows tested.

Value

An atlas_constraint.

Details

Monotonicity is non-strict, so a model that does not use var at all passes trivially (a flat response is monotone). This makes the constraint conditional - "if the model responds to var, the effect must be monotone" - without forcing the variable in. Pair with con_uses() when the variable must also be used.

Examples

mono <- con_monotone("wt", "decreasing")
mono$check(lm(mpg ~ wt, mtcars), mtcars)          # TRUE: linear, negative
#> [1] TRUE
con_monotone("wt", "increasing")$check(lm(mpg ~ wt, mtcars), mtcars)
#> [1] "predictions are not monotonically increasing in `wt` (violated with other predictors held at row 1's values)"

if (FALSE) { # \dontrun{
atlas(df, "price", constraints = list(
  mono_sqft = con_monotone("sqft", "increasing")
))
} # }