Skip to content

Squared

Squared loss, also known as the L2 loss.

Mathematically, it is defined as

\[L = (p_i - y_i) ^ 2\]

It's gradient w.r.t. to \(p_i\) is

\[\frac{\partial L}{\partial p_i} = 2 (p_i - y_i)\]

One thing to note is that this convention is consistent with Vowpal Wabbit and PyTorch, but not with scikit-learn. Indeed, scikit-learn divides the loss by 2, making the 2 disappear in the gradient.

Examples

from river import optim

loss = optim.losses.Squared()
loss(-4, 5)
81
loss.gradient(-4, 5)
18
loss.gradient(5, -4)
-18

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).