Weight by Information Gain Ratio

yzan

Operator Weight by Information Gain Ratio returns an unhelpful error message:

Process Failed<br>The setup does not seem to contain any obvious error, but...<br>

when one of the input attributes contains only missing values.

Proposed solution: Return 0 for attributes without any data.

Scope of the issue: The same issue applies to all weighting operators (Chi2 fails as well...).

Workaround: Use Remove Useless Attributes before feature weighting.

sgenzer

hi @yzan looks like a feature request as you're pointing out an error bubble that does work but clearly could be more helpful. Can you pls share an XML so I can replicate?

yzan

If you take the approach of returning an error (which is a valid solution), consider implementation of "realtime problem inspection", if it is technically possible at all. Reasoning: Unexpected process termination in the middle of a long run is unpleasing.

Details: The issue was detected in RapidMiner 9.4 (where e.g.: Gini index returns java.lang.ArrayIndexOutOfBoundsException). In RapidMiner 9.5, it just returns "SEVERE: null".

sgenzer

ah thank you. When I run that I do not get a red bubble - just a nasty "this process did not run" and a null pointer exception as you say. That's a bug.

So there is some realtime problem inspection that has been in RapidMiner for, well, forever. But there is always room for improvement!

Scott

IngoRM

If you take the approach of returning an error (which is a valid solution), consider implementation of "realtime problem inspection", if it is technically possible at all.

By the way, to handle this kind of problem you can also put the (potentially) problematic operator(s) inside of a Handle Exception operator.  In this case the process termination in the middle of a long process would at least be avoided...

Hope this helps,
Ingo