I have tried to use partitioning with decision tree (like this example: https://www.knime.org/nodeguide/analytics/classification-and-predictive-modelling/example-for-learning-a-decision-tree)
But in Decision tree predictor is reported this error "spec must not be null", and I don't understand the reason.
Thank you in advance!
this looks like you stumbled over a bug.
Could you send me a workflow to recreate the problem? Which version of the KNIME Analytics Platform are you currently using?
Thank you and best regards, Iris
Any news on this? I'm experiencing it as well, one year later.
- Suffix for probability columns checked throws error "spec must not be null"
In terms of reproducing the bug, I've never seen that checkbox work, yet. could it be related to dataset in some way?
Note, I'm working with category (non-numeric) data sets. I understand the knime nodes are based on C4.5. FYI, I ran the same dataset in R (with C5.0 which I read is a derivative of C4.5), and I can out probabilities there. The one difference I recall was that I had to do more data cleaning of the input strings (R/C5.0 errored out on any periods/fullstops in strings).
Also, Knime's C4.5 seems to perform slightly better, but that might just be optimized hidden params used by Knime with C4.5
Through trial and error testing multiple theories I managed to get probability output from the node.
Then I attempted to ascertain the cause, I now suspect the node has a limit of 60 classes. For a dozen test cases I tried, any time there were more than 60 classes it threw this error.
Is it possible to raise the class limit?
I have found the same thing in terms of the node limit. The thing is I can get the Decision Tree Learner to build the pmml model but it fails with the above error at scoring. I used a generic PMML predictor node however, and it worked. This suggests strongly that the Decision Tree Predictor has a bug in it.