Which metrics for multiple target in feedforward DL4j regression?

KNIME Extensions
1

Hello,
I attach below my workflow for a regression using feedforward dl4j regression.
With the workflow, I attach the 2 csv files for the features and the 12 outputs.
It is straightforward to do the prediction but I am not sure how to evaluate and to visualise the performance of the regression as the numeric scorer allows me to choose only one output.
Is there a node suitable to evaluate the multitarget problem?
Here is the link for the workflow and the csv files.

Thanks,

2

Hallo @zizoo,

unfortunately there is not designated node for this. You could either use the scorer node inside a loop if you want to evaluate every target separately or you could use the unpivot node and score all together.

Kind regards,
Janina

3

Hi @janina
I did not get the idea of using unpivot.
Fr the scorer in the loop, can I use the average of all the R2 for each output and its predicted column?
Thanks,

4

Hi Zizoo,

I have attached a modified workflow with Janina’s suggestion. The Pivoting could be complicated. That is why I used loops

  1. to collect all the actual and predicted values in two columns and get an aggregated over scores
  2. get scores on individual (Actual & Prediction) columns pairs

Hope this is helpful.

Best,
Temesgenmultioutput.knwf (3.8 MB)

1 Like
5

This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.