I understand that weka's J48 uses Gain Ratio as a method in determining how to split the tree, but when I compare the first attribute in the attribute ranking and the attribute by which my decision tree becomes split first, they are different (the first in rank in the attribute selector is the attribute "Domain", while the attribute by which the tree was split first was "MgO"). Why is it so? Aren't they supposed to be the same?
They are two different things--J48 applies the information gain locally at each node to select the best split for that node based on the data at that node. However, the use of Ranker and e.g., InfoGainAttributeEval evaluate the attributes globally based on all the data. On the other hand, rank of the attribute can only play a role when there is a tie in information gain between two different attributes where 48 breaks the ties with considering attributes order in the input dataset.
HTH.
Samer
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining