Optimal treshold value (optimal cut-off value) search

KNIME Analytics Platform
1

Hi,

For my model with binary classificator I need to have maximal value of sensitivity and specifity summ by choosing optimal cut-off ((treshold) value:

Can I do this calculatins in KNIME anyway?

Thank you,

Max

2

Not too sure what you are asking.

Is it normalising the data you are after with the Normaliser node.

Or removing datapoints above a certain value using the Row Filter node ?

 

Simon.

3

Hi, Simon,

I need to predict class in binary classification problem (MNN - NotMNN). In my case Logistic regression gives good result, so I have probability of MNN. If this probability more or equal 0.5 (cut-off value) - this is MNN class. In theory (http://www.basegroup.ru/library/analysis/regression/logistic on Russian, sorry, and here - http://epicentre.massey.ac.nz/resources/acvsc_grp/docs/Greiner_et_al_2000.pdf, part "3.2. Use of ROC analysis for the selection of cut-off values") I can change 0.5 to another value (0.4, 0.7), so sensitivity and specifity will change. I need to change this parameters in my case (medical problem, need more specific model). As I see - this is Linear Optimization problem. And I need to solve it in this case.

4

You may try to apply a transformation to the output of the regression (or any predictor) using the Math Formula node. By this you can transform the original values to your need, so that e.g. 0.4 is now 0.5.

5

Thank you, thor,

I do this, but can I find optimal value - not by hand?

6

The Optimization extension contains a parameter optimization loop. You can iterate over one or more parameters and it finds the combination where a certain function (e.g. error rate) is minimal or maximal. The public workflow server contains an example workflow (014001_ParameterOptimization).