Poisson

Poisson loss.

The Poisson loss is usually more suited for regression with count data than the squared loss.

Mathematically, it is defined as

$$L=exp(p_i)-y_i\times p_i$$

Its gradient w.r.t. to $p_i$ is

$$\frac{\partial L}{\partial p_i}=exp(p_i)-y_i$$

Methods

call

Returns the loss.

Parameters

• y_true
• y_pred

Returns

The loss(es).

gradient

Return the gradient with respect to y_pred.

Parameters

• y_true
• y_pred

Returns

The gradient(s).

mean_func

Mean function.

This is the inverse of the link function. Typically, a loss function takes as input the raw output of a model. In the case of classification, the raw output would be logits. The mean function can be used to convert the raw output into a value that makes sense to the user, such as a probability.

Parameters

• y_pred

Returns

The adjusted prediction(s).