Interpreting the results of SOM/Kohonen nodes

[genericuserid111]

Hi All,

I'm working on a creating a predictive model, and would like to explore the data (there's a LOT of it) before starting the actual modeling work. I went through one of the EM courses (Applied Analytics Using SAS Enterprise Miner) but found it didn't explain the interpretation of the SOM/Kohonen node.

This may just be because I'm new to EM and data analysis, and if so I'd appreciate if someone could point me in the direction of a resource to learn how to interpret SOM/Kohonen maps, but from my searching and reading (a few textbooks recommended for neural networks) I'm still unsure of the meaning of the nodes output.

Any help would be greatly appreciated!

Thank you

[M_Maldonado · Accepted solution]

In this article the authors use the Segment Profile node to interpret the segments that the SOM/K node outputs.

I hope this helps,

Miguel

http://www.iasri.res.in/sscnars/data_mining/8-som%20with%20e-miner.pdf

[genericuserid111]

Thank you for your reply!

Unfortunately, the article doesn't explain *what* the Segment Profile nodes results mean. It may just be that I'm unfamiliar with data analysis, but information about things like "Petal width" or "Som Dimension2" doesn't really mean anything to me.

I've found that most tutorials or explanations of EM simply show you how to perform a specific task without providing the information of why. Again though, it's very possible that I just am lacking the knowledge needed to make use of the tool.

[gergely_batho]

Hi,

Segment Profile Node creates different stats and graphs about the clusters or segments. It does nothing specific for SOM/Kohonen.

So, the interpretation of the Segment Profile node is the same as you would use it after a "regular" Cluster Node.

SOM/Kohonen has one additional feature compared to other methods: each cluster (which corresponds to an output neuron) has a row and a column number. Clusters are forced into a grid. Every cluster will have 2 or 4 neighbors...

You could visualize this grid (heat map?), maybe coloring based on number of obs, or average of some variable. It would be like a nonlinear 2D mapping of your higher dimensional data. (Remember: by default SOM initializes with a kind of linear 2D mapping based on the first 2 principal components.)

[genericuserid111]

It's seeming like I'm the limiting feature here as I have trouble interpreting the output for the cluster node as well.

Would the interpretation of the output from the cluster node be considered something a data analyst would understand out of the box? I.e. is there some standard of terminology SAS is adhering to here?

Thank you!