gbt.ksval {agtboost} | R Documentation |
gbt.ksval
transforms observations to U(0,1) if the model
is correct and performs a Kolmogorov-Smirnov test for uniformity.
gbt.ksval(object, y, x)
object |
Object or pointer to object of class |
y |
Observations to be tested |
x |
design matrix for training. Must be of type |
Model validation of model passed as object
using observations y
.
Assuming the loss is a negative log-likelihood and thus a probabilistic model,
the transformation
u = F_Y(y;x,\theta) \sim U(0,1),
is usually valid.
One parameter, \mu=g^{-1}(f(x))
, is given by the model. Remaining parameters
are estimated globally over feature space, assuming they are constant.
This then allow the above transformation to be exploited, so that the
Kolmogorov-Smirnov test for uniformity can be performed.
If the response is a count model (poisson
or negbinom
), the transformation
u_i = F_Y(y_i-1;x,\theta) + Uf_Y(y_i,x,\theta), ~ U \sim U(0,1)
is used to obtain a continuous transformation to the unit interval, which, if the model is correct, will give standard uniform random variables.
Kolmogorov-Smirnov test of model
## Gaussian regression:
x_tr <- as.matrix(runif(500, 0, 4))
y_tr <- rnorm(500, x_tr, 1)
x_te <- as.matrix(runif(500, 0, 4))
y_te <- rnorm(500, x_te, 1)
mod <- gbt.train(y_tr, x_tr)
gbt.ksval(mod, y_te, x_te)