RegularizedMixtureNLLLoss

class mlquantify.losses.RegularizedMixtureNLLLoss(tau_0=0.0, tau_1=0.0, reduction='mean')

Mixture NLL with optional ordinal-smoothness regularization.

Extends MixtureNegativeLogLikelihoodLoss with first- and second-order penalty terms that encourage a smooth prevalence vector when the classes have an ordinal interpretation.

The total loss is

\[\mathcal{L} = \text{NLL}(\hat{p}) + \frac{\tau_0}{2} \sum_{c} (\hat{p}_{c+1} - \hat{p}_c)^2 + \frac{\tau_1}{2} \sum_{c} (-\hat{p}_{c} + 2\hat{p}_{c+1} - \hat{p}_{c+2})^2\]

Parameters:

tau_0float, default=0.0

Weight of the first-order smoothness penalty. Set to 0 to disable.

tau_1float, default=0.0

Weight of the second-order smoothness penalty. Set to 0 to disable.

reduction{‘mean’, ‘sum’}, default=’mean’

How to reduce the per-sample log-likelihood values.

Attributes:

tau_0float

First-order regularization weight.

tau_1float

Second-order regularization weight.

reductionstr

The configured reduction mode.

Examples

>>> from mlquantify.losses import get_loss
>>> import numpy as np
>>> loss = get_loss("regularized_mixture_nll", tau_0=0.1)
>>> prev = np.array([0.3, 0.4, 0.3])
>>> lkl = np.array([[0.2, 0.5, 0.3], [0.6, 0.3, 0.1], [0.1, 0.2, 0.7]])
>>> loss(prev, lkl)  
1.1...

__call__(prevalences, class_likelihoods)

Compute the regularized mixture negative log-likelihood.

Parameters:

prevalencesarray-like of shape (n_classes,)

Estimated class prevalence vector.

class_likelihoodsarray-like of shape (n_classes, n_samples) or (n_samples, n_classes)

Per-class likelihood for each test instance.

Returns:

lossfloat

Mixture NLL plus any active smoothness penalties.

Examples

>>> from mlquantify.losses import get_loss
>>> import numpy as np
>>> loss = get_loss("regularized_ml", tau_0=0.0)
>>> prev = np.array([0.5, 0.5])
>>> lkl = np.array([[0.2, 0.8], [0.7, 0.3]])
>>> round(loss(prev, lkl), 4)
0.6931