Fits the baseline SPQR (Spline-based Probabilistic Quantile Regression) model using a neural network parameterization of spline mixture weights.
Usage
fit.spqr(
input_dim,
hidden_dim,
n.knots,
x_training,
x_validation,
y_training,
y_validation,
hyperparameter = NULL,
pre_normalize = FALSE,
package.it = TRUE
)Arguments
- input_dim
Integer. Number of covariates (input features).
Integer vector specifying hidden layer sizes.
- n.knots
Integer. Total number of spline basis functions.
- x_training
Numeric matrix of training covariates.
- x_validation
Numeric matrix of validation covariates.
- y_training
Numeric vector or matrix of training responses.
- y_validation
Numeric vector or matrix of validation responses.
- hyperparameter
List created by
create.packages.hyperparameter()specifying training configuration (e.g., epochs, batch size).- pre_normalize
Logical. Included for API consistency. Currently does not alter preprocessing behavior.
- package.it
Logical. If TRUE (default), returns a structured
"spqrx_model"object. If FALSE, returns the rawkeras_model.
Value
If package.it = TRUE, returns an object of class
"spqrx_model" containing:
The trained keras model
Spline knot locations
Normalization parameters
Variable names (if available)
Otherwise, returns a trained keras_model object.
Details
This function performs:
Response rescaling to the unit interval
Covariate standardization using combined training/validation statistics
Construction of spline basis and integrated basis matrices
Neural network training via
in.fit.spqr()
The fitted model can optionally be returned as a packaged
"spqrx_model" object containing normalization metadata
and spline information for downstream prediction and explainability.
The response variable is internally rescaled to the unit interval prior to spline construction. Covariates are standardized using mean and standard deviation computed from the combined training and validation data.