Opened 9 months ago

Last modified 18 hours ago

#2910 readytorelease defect

ImpactCalculation for IDiscriminantClassificationSolutions is wrong

Reported by: mkommend Owned by: mkommend
Priority: medium Milestone: HeuristicLab 3.3.16
Component: Problems.DataAnalysis Version: trunk
Keywords: Cc:

Description (last modified by mkommend)

The reason therefore is that the model parameters (threshold and class values) are not updated during modification of the model. As a result suboptimal node and variable impacts are calculated.

Change History (6)

comment:1 Changed 9 months ago by mkommend

  • Status changed from new to accepted
  • Version set to trunk

comment:2 Changed 9 months ago by mkommend

  • Description modified (diff)

comment:3 Changed 9 months ago by mkommend

  • Owner changed from mkommend to abeham
  • Status changed from accepted to reviewing

r15871: Added recalculation of thresholds for IDiscriminantClassificationModels during impact calculation.

I tested these changes with an existing SymbolicClassificationModel. However, I am unsure about the side effects of cloning the models that are introduced (GaussianProcesses, GradientBoostedTrees, ...).

comment:4 Changed 3 months ago by abeham

  • Owner changed from abeham to mkommend
  • Status changed from reviewing to assigned

I noticed that in the Simplifier when you move the threshold line, it constantly updates the UI on the model side and recalculates parameters. This may probably be somewhat expensive. I don't know if this is a problem.

comment:5 Changed 2 months ago by mkommend

  • Status changed from assigned to readytorelease

Thank you for reporting this issue. It should not pose a problem, because the calculation is aborted (AFAIK), when the model is updated in the mean time.

comment:6 Changed 18 hours ago by gkronber

Tried to merge r15871 but this fails.

#2884 and #2871 also have changes on ClassificationSolutionVariableImpactsCalculator and must be merged first.

Note: See TracTickets for help on using tickets.