Problems with PMML denormalizer

KNIME Analytics Platform
1

Hi all,

I'm new to Knime and I'm having some trouble to get denormalized values of both Regression Predictor and MultiLayerPerceptron Predictor.

First  I put the whole dataset through the normalizer node, then I partitioned it into training and validation sets. After obtaining  the linear and NN models, I send them to Regression Predictor and MultiLayerPerceptron Predictor, along with the validation data set. Both predictor results are sent to denormalizer nodes. The problem is that when I look into the denormalizer outputs, the predicted values are still normalized. Looks like I should provide the denormalization info of the new predicted variables  in order to get it done, but I can't figure how and could't find any examples  compatible to 3.4.1 version. I'll be grateful for any help.

2

Hi,

and welcome to the KNIME Community.

To denormalize the prediction the name of the column with the predictions has to be the same as the column name of the target value in the normalizer node. We can therefore define a customized name for the prediction column in the configuration window to the "Predictor node".

But we have to have in mind that the column names are unique and KNIME will add a (#1) if we have already a column with this name. This means that in case of a test set where the target variable is available we would have to use the denormalizer node twice.

Did this help already? Otherwise I can provide you also an example.

Cheers

Kathrin

3

Hey,

bit late, but I had the same problem some day.
If you realy have to normalize the dependent variable then I see no other way than the one Kathrin explained - at least in Knime.

I assume that the key is, that you may not have to normalize the dependent variable at all.
Normalizing the independent variables transforms them all in a comparable scale, to prevent that one variable dominates the model just because of its scale.
That is not a problem when you look at the dependent variable.
I’m not a professional in this field, but I think you can just let your dependent variable be as it is - then you also have no problems with normalized predictions.
Hope this helps…

1 Like