I currently use several methods for the analysis of the same dataset. I want to observe identical rules obtained by several classifiers. However, the output depends on the classification method, thus how to know if the obtained rules/tree contain the same information?
Example:
I want an automatic way to find that this rule :
A = 1 AND B = 0: 1
is identical to this tree:
A = 0: 0
A = 1
| B = 0: 1
| B = 1: 0
Thank you for your answer. However the question is not identical performance but identical rules. Indeed, I have already selected classifiers with high AUC obtained with several algorithms. However, I want to reduce the number of classifiers by picking just one of them when rules or trees contain exactly the same information.
In the example the first classifier obtained by a first method is " if A equal 1 and B equal zero, then class 1"
This classifier is obtained also by another method (tree based method) but the presentation of the result is not the same :
A = 0: 0
A = 1
| B = 0: 1
| B = 1: 0
As I have hundred of classifiers to compare I was wondering if a method already exist in order to obtain this kind of information.
Two approaches: The first one is to convert the decision tree to a set of decision rules and compare the two rule sets. In your example the tree will give 3 rules: A = 0: 0 , A = 1 AND B = 0: 1, A = 1 AND B = 1: 0. From this set, remove the rules produced by the alternative method, and see if all are removed, or if the remaining rules predict a default class value. If this happens, then the rule set and the tree are equivalent.
The second approach is a probabilistic one: Generate test data sets and feed them to the rules and the tree. If they produce the same results, then they are probably equivalent. Of course if the domain is small enough to allow a data set to be exhaustive, then producing the same output means the rules and the tree are definitely equivalent.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs