perfsmooth: Find an optimal monotonic performance curve.

In rplzzz/perfsmooth: Fit smooth, monotonic curve to performance data

Description

Find a curve that best matches a set of points, subject to the constraint that the curve must be monotonic.

Usage

perfsmooth(x, y, N, maxdegree = 5, xinterval = c(0, 1), itmax = 2500)

Arguments

x	x-values of control points.
y	y-values of control points. These should be fractions between 0 and 1.
N	Number of cases that contributed to each y value.
maxdegree	Maximum degree to use in the Legendre series. Must be <= 10.
xinterval	Domain of valid values for x (whether or not the entire interval is actually represented in the x values). Default is [0,1].
itmax	Maximum number of iterations in the solver

Details

The points are assumed to be observations of a binary event, which give an estimate of the probability of the event. The uncertainty of those probability values is therefore beta distributed with parameters derived from the number of observations.

The intended use here is for a case where the x-values are a measure of algorithm sensitivity (with 0 being highly selective and 1 being highly sensitive), and the y-values are PPV (or similar measure). A more selective algorithm should generate fewer false positives, but with fewer cases, there is more uncertainty in the measured value.

The smoothed curve is represented as a series of Legendre polynomials, up to degree 10 (or less, if specified). The return value is the vector of series coefficients.

Value

Vector of Legendre series coefficients.

rplzzz/perfsmooth documentation built on Dec. 22, 2021, 7:14 p.m.