What does "splitter" attribute in sklearn's DecisionTreeClassifier do?

Question

The sklearn DecisionTreeClassifier has a attribute called "splitter" , it is set to "best" by default, what does setting it to "best" or "random" do? I couldn't find enough information from the official documentation.

Answer 1

Short ans:

RandomSplitter initiates a **random split on each chosen feature**, whereas BestSplitter goes through **all possible splits on each chosen feature**.

---

Longer explanation:

This is clear when you go thru _splitter.pyx.

RandomSplitter calculates improvement only on threshold that is randomly initiated (ref. lines 761 and 801). BestSplitter goes through all possible splits in a while loop (ref. lines 436 (which is where loop starts) and 462). [Note: Lines are in relation to version 0.21.2.]

As opposed to earlier responses from 15 Oct 2017 and 1 Feb 2018, RandomSplitter and BestSplitter both loop through all relevant features. This is also evident in _splitter.pyx.

Answer 2

In fact, the "random" parameter is used for implementing the extra randomized tree in sklearn. In a nutshell, this parameter means that the splitting algorithm will traverse all features but only randomly choose the splitting point between the maximum feature value and the minimum feature value. If you are interested in the algorithm's details, you can refer to this paper [1]. Moreover, if you are interested in the detailed implementation of this algorithm, you can refer to this page.

[1]. P. Geurts, D. Ernst., and L. Wehenkel, “Extremely randomized trees”, Machine Learning, 63(1), 3-42, 2006.