Mdl = gamtrain (X, Y, Knots, Order, Method, Inter, P1, P2) fits one univariate spline per column of X and returns the additive model as a structure. It is the fitting engine shared by ClassificationGAM and RegressionGAM, and it is not meant to be called directly.

• X is an $NxP$ numeric matrix of predictors. A missing value is not an error: the observation is left out of the affected predictor’s spline and its prediction from that term is NaN.
• Y is an $Nx1$ numeric vector of responses. For Method 1 it must hold zeros and ones.
• Knots is a $1xP$ vector giving the number of spline pieces for each predictor, and Order a $1xP$ vector giving the degree of the polynomial on each piece. A spline of $K$ pieces and degree $D$ spans a space of $K\; +\; D$ dimensions.
• Method selects the fitting scheme: 1 boosts the log-odds by gradient descent, as a classifier is fitted, and 2 backfits the partial residuals, as a regression is fitted.
• Inter is the intercept the fit starts from: a proportion for Method 1, which is stored as its log-odds, and the response mean for Method 2. A proportion of zero or one is not an error: its log-odds is infinite, the gradient is zero throughout and every additive term stays at zero, which is the fit a single-class response has.
• P1 and P2 are the scheme’s two parameters. For Method 1 they are the learning rate and the number of boosting iterations; for Method 2 the convergence tolerance and the maximum number of backfitting cycles.

Mdl is a structure with the following fields.

• Intercept, the constant term the additive terms are added to.
• Parameters, a $1xP$ structure array of piecewise polynomials in the form ppval consumes, one per predictor.
• Iterations, the number of iterations performed.
• Residuals, the $Nx1$ residual vector at the last iteration.
• RSS, the scalar residual sum of squares for Method 1 and the $1xP$ per-term criterion the backfitting stops on for Method 2.

See also: gampredict, ClassificationGAM, RegressionGAM, fitcgam, fitrgam

Source Code: gamtrain