Cannot genereate ROC from HBOS example set

MaartenK
MaartenK New Altair Community Member
edited November 5 in Community Q&A
Hello,

I have a model using HBOS to calculate an outlier score. I want to feed that into an "GenerateROC" component to evauluate the scores, like with other outlier detection components.
However, the GenerateROC component gives an design time error that the exampleset coming from the HBOS in facts is not an example set. I use the output connector at the top, without the colors. 
I also tried saving the example set first and the feed it in the GenerateROC component in a new process. 
In that case, the GenerateROC does give a run time error saying "' '' not found in the labels".
There seems to be something wrong with the metadata of the HBOS output example set. Any help with be great.

Kind regards,

Maarten

Best Answer

  • Telcontar120
    Telcontar120 New Altair Community Member
    Answer ✓
    There is indeed a known problem with the metadata from the HBOS operator.  Unfortunately it is a third party extension so RapidMiner has no direct control over that.
    Having said that, you can get it to work.  Simply use the Store operator after the HBOS operator, and then Retrieve the exampleset and it will have the appropriate metadata.  However, to generate an ROC curve, you will need to do some additional work.  You will need to use the HBOS score to generate a prediction and a confidence, set those roles accordingly, and then you will also need to set a label by creating an outlier/non-outlier flag (manually or with some other logic). 

Answers

  • Telcontar120
    Telcontar120 New Altair Community Member
    Answer ✓
    There is indeed a known problem with the metadata from the HBOS operator.  Unfortunately it is a third party extension so RapidMiner has no direct control over that.
    Having said that, you can get it to work.  Simply use the Store operator after the HBOS operator, and then Retrieve the exampleset and it will have the appropriate metadata.  However, to generate an ROC curve, you will need to do some additional work.  You will need to use the HBOS score to generate a prediction and a confidence, set those roles accordingly, and then you will also need to set a label by creating an outlier/non-outlier flag (manually or with some other logic).