penalized_negbin
Regularied negative-binomial marginal model.
PenalizedNegBin(formula, *, penalties=(), device=None)
:class:NegBin with penalty P added to the training loss
This is a regularized form of the NegBin marginal model. By default, this
does not use Poisson IRLS during initialization, since regularized settings
typically have high-dimensional covariates, which are unstable in an
ordinary linear model. These initial Poisson IRLS steps can be added back by
passing max_iter to fit.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
formula
|
dict of str to str, or str
|
|
required |
penalties
|
sequence of CoefficientPenalty
|
See :mod: |
()
|
device
|
device or str
|
|
None
|
Examples:
>>> from scdesigner.distributions.penalties import CoefficientPenalty, QuadraticPenalty
>>> penalty = CoefficientPenalty("mean", ["x"], [1.0], QuadraticPenalty(lam=0.1))
>>> model = PenalizedNegBin({"mean": "~ x", "dispersion": "~ 1"}, penalties=[penalty])
>>> type(model).__name__
'NegBin'
Source code in scdesigner/src/scdesigner/distributions/penalized_negbin.py
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 | |