Fit a conformal anomaly detector from a Tier R healthy reference cohort

Fits a conformal anomaly detector using k-nearest-neighbour distance as the non-conformity measure. A held-out calibration fraction of the healthy reference cohort is used to calibrate the conformal p-value, providing a distribution-free, finite-sample FPR guarantee at level alpha (Vovk et al., 2005; Romano et al., 2019).

Usage

fit_conformal_anomaly(
  X_ref,
  representation = identity,
  k = 10L,
  calibration_split = 0.3,
  seed = 42L
)

Arguments

X_ref

Numeric matrix (samples \(\times\) features) of healthy reference samples.

representation

Function to project samples to a latent space. Default identity (use raw features).

k

Integer - number of nearest neighbours. Default 10.

calibration_split

Fraction of X_ref held out for calibration. Default 0.30.

seed

Integer - RNG seed for the train/calibration split. Default 42.

Value

Object of class conformal_anomaly_fit.

References

Vovk V, Gammerman A, Shafer G. (2005) Algorithmic Learning in a Random World. Springer.

Examples

if (FALSE) { # \dontrun{
set.seed(42)
X_ref <- matrix(rnorm(200 * 30), nrow = 200)
fit <- fit_conformal_anomaly(X_ref, k = 10L)
fit$n_cal
} # }