Absolute¶
Absolute loss, also known as the mean absolute error or L1 loss.
Mathematically, it is defined as
\[L = |p_i - y_i|\]
Its gradient w.r.t. to \(p_i\) is
\[\frac{\partial L}{\partial p_i} = sgn(p_i - y_i)\]
Examples¶
from river import optim
loss = optim.losses.Absolute()
loss(-42, 42)
84
loss.gradient(1, 2)
1
loss.gradient(2, 1)
-1
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).