Opened 9 months ago

Last modified 4 months ago

#2660 accepted enhancement

Extend VariableNetworkInstances to include problems with no noise and to calculate actual variable relevance

Reported by: bburlacu Owned by: gkronber
Priority: medium Milestone: HeuristicLab 3.3.15
Component: Problems.Instances Version: 3.3.14
Keywords: Cc:

Description

It should also be possible to use VariableNetwork problem instances with zero noise. Additionally, the problem descriptions could be improved such that variable descriptions are ordered alphabetically.

Change History (11)

comment:1 Changed 9 months ago by bburlacu

  • Owner set to bburlacu
  • Status changed from new to accepted

r14260: Include instances with no noise. Improve formatting of problem description.

comment:2 Changed 9 months ago by bburlacu

  • Owner changed from bburlacu to gkronber
  • Status changed from accepted to reviewing

comment:3 Changed 9 months ago by gkronber

r14271: slightly refactored VariableNetwork instance and added calculation of variable relevance

comment:4 Changed 8 months ago by gkronber

  • Status changed from reviewing to assigned
  • Summary changed from Extend VariableNetworkInstances to include problems with no noise to Extend VariableNetworkInstances to include problems with no noise and to calculate actual variable relevance

comment:5 Changed 8 months ago by gkronber

The variable relevance measure introduced with r14271 does not completely match our definition of variable relevance. The variable relevance calculation should therefore be adapted.

comment:6 Changed 8 months ago by gkronber

r14291: changed calculation of variable relevance values for variable interaction networks based on sampling Gaussian processes.

Instead of taking inverse length scale (ARD). We calculate the deviation from the original function (y) after individual variables are removed (y').

comment:7 Changed 8 months ago by gkronber

  • Owner changed from gkronber to bburlacu
  • Status changed from assigned to reviewing

comment:8 follow-up: Changed 8 months ago by gkronber

  • Owner changed from bburlacu to gkronber
  • Status changed from reviewing to assigned

The current variable relevance calculation does not consider relevances of indirectly dependent variables.

comment:9 Changed 4 months ago by gkronber

  • Status changed from assigned to accepted

comment:10 in reply to: ↑ 8 Changed 4 months ago by gkronber

Replying to gkronber:

The current variable relevance calculation does not consider relevances of indirectly dependent variables.

This is in fact the correct approach.

When we sample data form the network we also observe statistical dependencies to other - indirectly related - variables. However, these indirect dependencies should not be considered as actual dependencies within the network (indirectly connected variables are conditionally independent).

Last edited 4 months ago by gkronber (previous) (diff)

comment:11 Changed 4 months ago by gkronber

r14623: removed obsolete comment

Note: See TracTickets for help on using tickets.