I am predicting a target $y$ using regression. In my application, the prediction $\hat{y}$ should be always no less than $y$. If $y>\hat{y}$, it is definitely a wrong prediction. On the $y<\hat{y}$ side, I want $\hat{y}$ is close to $y$.

In this case, I cannot use $(\hat{y}-y)^2$ to measure the error because it measures two sides. I guess it should be a common case in ML practices. Is there any commonly used continuous loss function that measures on-side error?


How about something like: \begin{eqnarray} (y-\hat{y})^2 , y < \hat{y} \\ (y-\hat{y})^3 , y \geq \hat{y} \end{eqnarray}

This is both continuous and differentiable, and penalizes $y > \hat{y}$ more severely than $y < \hat{y}$. You can even control the severity of the penalization by changing the exponent from 3 to a higher odd number.

Here's how it looks ($x = y - \hat{y}$):

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.