# What is the loss/cost function of decision trees?

In Decision Tree, splitting criterion methods are applied say information gain to split the current tree node to built a decision tree, but in many machine learning problems, normally there is a cost/loss function to be minimised to get the best parameters.

My question is how to define such a cost function of Decision Tree?

• For some outcome $y$, decision trees will give you predictions $\hat{y}$. You may then choose the tree that has the minimum squared error, which means you're working with the typical loss function $L=(y-\hat{y})^2$. – Alvaro Fuentes Feb 27 '18 at 14:06