I am trying to use a parameter optimization loop (OL) to get the best parameters for the SVM Learner node.
Within the OL I perform 10x CV (Counting loop). I turn the 10 PMML models into a PMML Ensemble Model.
My problem is to aggregate the PMML ensemble models such that I can choose the model which was created with the best parameters.
Using the Table creator node I couldn’t add PMML cells.
The current version of the workflow is on KNIME Hub: WF on KNIME Hub
Update …
I created a table with the expected signature and used that to add the PMML models.
Unfortunately, only the last round of optimization was available in the final table (somehow concatenation didn’t work?).
I finally wrote each ensemble modle to a csv file with append mode and selected from the file.
Would this be possible without the detour via writing - reading files?
ML models are trained and tested on a subset of the data and that may lead to outliers. I am interested in the average behavior of the models for each optimization attempt. Therefore, I perform cross-validation with the counting loop (10-fold in this case). A side effect is that I can build an ensemble model from all models trained in the counting loop which may better represent the power of the optimization attempt.
ok thanks.
There is a X Partitioner and X Aggregator node in Knime to do cross validation.
In your workflow I did not see that the partition changes for each loop so i was confused.
bR