Graduation settings¶
Finally, the deployed tables are smoothed with Whittaker-Henderson graduation, the actuarial
smoother for rating factors. Each table picks its own strength by generalized cross-validation
(GCV), so a table whose roughness is real shape is left untouched. No objects are held out for
it. The outcome is recorded in the graduation_report_ attribute. See
Pruning, banding and graduation for details.
Graduation is skipped for models with monotone_constraints and is not supported for
multiclassification.
graduate¶
Description¶
Smooth the deployed tables with Whittaker-Henderson graduation.
None and True turn graduation on, and False turns it off (the graduation_report_
attribute is then not set). An explicit True requires prune=True, and raises an error for
multiclassification. Whether a smoothed table is what you want to deploy is a modelling
decision: graduation has no accuracy gate.
Type
bool
Default value
None (on)
graduation_alpha¶
Description¶
A fixed smoothing strength for every table, instead of the strength each table picks by
generalized cross-validation. 0 turns off the smoothing of the dense tables, which is useful
together with graduation_high_order_alpha.
Type
float
Default value
None (each table picks its own strength)
graduation_high_order_alpha¶
Description¶
The strength, in the range \([0; 1]\), of an additional neighbour smoothing step for effects of order 3 to 8 that are stored in factored form. Unlike the smoothing of the dense tables, its strength is fixed rather than selected by cross-validation.
Numeric axes are smoothed along their order. Categorical axes and the edges between missing and
non-missing values are not smoothed. Effects with more than 1 024 stored boxes, or that would
add more than 4 096 boxes in total or 2 000 000 cells, are skipped; the details are in
graduation_report_.
0 turns it off. Requires prune=True and graduation turned on; not supported with
monotone_constraints or for multiclassification.
Type
float
Default value
0.0 (off)