The Altair Community is migrating to a new platform to provide a better experience for you. The RapidMiner Community will merge with the Altair Community at the same time. In preparation for the migration, both communities are on read-only mode from July 15th - July 24th, 2024. Technical support via cases will continue to work as is. For any urgent requests from Students/Faculty members, please submit the form linked here.
Options

"How to Validatte SOM Clusters?"

ppsheehyppsheehy Member Posts: 5 Contributor II
edited June 2019 in Help
Hi,

What’s the best way to validate the output of SOM clusters? I tried using ‘Apply Model’ & ‘%Performance’ directly on the Pre Processing model, but that doesn’t work.

I then used a Neural Net on the output of the SOM and Added ‘Apply Model’ and ‘%Performance’ on the output of that. It seems to work (I have manually checked some of the Accuracy values). However, I am worried that by adding in the Neural Net, I am now measuring how well the Neural Net is performing, not the SOM…

Is there a better way to check the accuracy of the clusters a SOM creates?

Thanks in advance,

Paul S
Tagged:

Answers

  • Options
    landland RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 2,531 Unicorn
    Hi,
    the som algorithm is for reducing the dimensionality in the first place. If you want to regard the position in the som node grid as a cluster, then you have to manually convert the attributes generated by the som into something identifying a cluster.
    If you use a 1-dimensional SOM, simply convert the numerical attribute into a polynomial one (numerical to polynomial) and set the role of the label to "cluster" using the Set Role operator. Then you might use the Map Cluster to Label operator for measuring the accuracy of the best possible fit.
    If you use more dimensions, you will have to generate an attribute containing the combination of all som attributes for identifying the cluster. Refer to Generate Attributes for that.

    Greetings,
      Sebastian
Sign In or Register to comment.