1
$\begingroup$

I read in a blog that the decision tree has this disadvantage:

Not fit for continuous variables

If this is true, then why?

$\endgroup$

4 Answers 4

1
$\begingroup$

Continuos Variable in training data (X) If you look at decision trees they try to split data based on categories in case of categorical data and based on threholds in case of continuous data. Now to split on continuous data it randomly tries to create BINs and calculate the entropy/ Gain etc whatever you have chosen. This implies that Decision Trees is able to use contnuous variable for training and their is no disadvanatge except it can take a little bit more time to train due to preocess of Bin creation and finding best split on the Bin.

Continuos Variable in target (X) Decision trees works well for regression problem but now objective is not to find splits which give maximum information gain but which reduces variance amongs onservation withing the same leaf nodes.

Based on my understanding, In general Decison Trees has been modified to work well with both continuous and categorical data

$\endgroup$
0
$\begingroup$

Decision trees work well with categorical variables because of the node structure of a tree. A categorical variable can be easily split at a node. For example, yes or no or 0 or 1. A continuous variable is just that, continuous along a range which cannot be easily split at a node. For example, a floating point number between 0 and 100.

$\endgroup$
1
  • $\begingroup$ Also the ordering of the categorical variables value matters to some extent also $\endgroup$
    – Aditya
    Commented May 11, 2018 at 1:43
0
$\begingroup$

It can't be used for micro trend detection as each node looks for values greater than or smaller than the threshold value.

$\endgroup$
0
$\begingroup$

Decision trees can handle continuous variables, so this shouldn't be a disadvantage. They just make it categorical by defining thresholds to separate continuous variables into classes.

However, if you have a model with continuous variables only, neural networks will probably be the better solution for such an use-case.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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