Mean Tweedie Deviance

mean-tweedie-deviance

Metric mean_tweedie_deviance from scikit-learn (sklearn.metrics.mean_tweedie_deviance)

When to invoke this skill

The user has predictions + ground truth and asks to evaluate with mean_tweedie_deviance, or mentions sklearn.metrics.mean_tweedie_deviance directly, or wants the standard scikit-learn implementation.

Reference signature

from sklearn.metrics import mean_tweedie_deviance

# mean_tweedie_deviance(y_true, y_pred, *, sample_weight=None, power=0)

Library docstring

Mean Tweedie deviance regression loss.

Read more in the :ref:`User Guide <mean_tweedie_deviance>`.

Parameters
----------
y_true : array-like of shape (n_samples,)
    Ground truth (correct) target values.

y_pred : array-like of shape (n_samples,)
    Estimated target values.

sample_weight : array-like of shape (n_samples,), default=None
    Sample weights.

power : float, default=0
    Tweedie power parameter. Either power <= 0 or power >= 1.

    The higher `p` the less weight is given to extreme
    deviations between true and predicted targets.

    - power < 0: Extreme stable distribution. Requires: y_pred > 0.
    - power = 0 : Normal distribution, output corresponds to
      mean_squared_error. y_true and y_pred can be any real numbers.
    - power = 1 : Poisson distribution. Requires: y_true >= 0 and
      y_pred > 0.
    - 1 < p < 2 : Compound Poisson distribution. Requires: y_true >= 0
      and y_pred > 0.
    - power = 2 : Gamma distribution. Requires: y_true > 0 and y_pred > 0.
    - power = 3 : Inverse Gaussian distribution. Requires: y_true > 0
      and y_pred > 0.
    - otherwise : Positive stable distribution. Requires: y_true > 0
      and y_pred > 0.

Returns
-------
loss : float
    A non-negative floating point value (the best value is 0.0).

Examples
--------
>>> from sklearn.metrics import mean_tweedie_deviance
>>> y_true = [2, 0, 1, 4]
>>> y_pred = [0.5, 0.5, 2., 2.]
>>> mean_tweedie_deviance(y_true, y_pred, power=1)
1.4260...

Quick recipe

import sklearn.metrics as _m
score = _m.mean_tweedie_deviance(y_true, y_pred)

Don'ts

• Don't reimplement when the library version handles edge cases (NaN, ties, empty inputs) better than a hand-rolled formula.
• Always check the library version's argument order — sklearn is (y_true, y_pred) while torchmetrics is (preds, target).