I was wondering what is the PMML version of extracted models of KNIME. From other PMML models (ver 3.0) i know that every document should have a node however such a node does not exist in the PMML extracted model of a KNIME neural network.
I also tried having an extracted PMML model read into KNIME and i get this exception :
ERROR Model Reader Execute failed
ERROR Model Reader line=2: Document root element "PMML", must match DOCTYPE root "null".
xml: URI=java.io.BufferedInputStream@1b66175
dtd: URI=null
I am certain that the PMML model to be read by Knime does have node on second line.
You are right, the PMML import/export is a bit odd - it is an XML based model representation but we have not yet migrated it to an official PMML standard. Also because most of the models in KNIME aren't expressable as PMML.
We'll put that on our to-do list for the obvious candidates at least.
Better PMML support will become available in 2.0 - which will show up on this site before Christmas. KNIME will then read/write standard PMML for the more commly used models.
We would appreciate help with the R-integration, however, since we had problems when trying to have KNIME R nodes create PMML output / accept PMML input.
I want to add some details about the upcoming PMML support in KNIME 2.0:
PMML models of versions 3.0, 3.1 and 3.2 can be imported. Models are exported as PMML v.3.1 only, i.e. the older models are converted if possible. Thus, KNIME will act as a PMML Producer and Consumer.
We started to support this functionality with the PMML core features only, extensions and x- marked attributes are ignored (as defined in the PMML standard). All the missing value treatment, statistics and distribution information are also ignored, since this can be modeled in KNIME with the help of additional nodes within the workflow.
The following nodes will output their learned model as PMML 3.1 (with the respective predictor nodes which are able to read PMML models):
Clustering (k-Means, fuzzy c-Means)
Neural Network
Support Vector Machine
Decision Tree
Linear and Polynominal Regression
I generated a PMML 4.2 with spss modeler (bagged neural network) and when I try to import it with KNIME I get this error:
ERROR PMMLValidator <LiftData targetFieldDisplayValue="1" targetFieldV[...]: Expected element 'ROC' instead of 'LiftData' here in element PredictiveModelQuality