408590_en-US

liliconcom

Grouping by fixed number(i.e. 4 or 5) of products in cluster.

Hi,
I got some question from some customer, I have tried to find solution but couldn't find until yet.
Hope your idea sharing.

Here is the situation.
They produce product B and measure 5 points in every products(lots) like below.
And, they want to group 5 most similar products based on the value of 5 points
(Assume there are 100 products, 5 points, and 500 measurement data)

1. We can group based on some statistics(mean, stdev and other methods to calculate flatness etc.) for 5 points.
But I think that these methods are less reasonable compared to Clustering.

2. I tried Hierarchy and K-Means. In this case, these methods has also some limitations.
   They want to group by fixed number(i.e. 4 or 5) of products in every cluster.
  I saw something similar about the grading of agricultural products, but I couldn’t remember.

 

Is there any other ideas ?
Thanks in advance
Ikju

 

Re: Grouping by fixed number(i.e. 4 or 5) of products in cluster.

Ikju: What’s the limitation of clustering? You didn’t get the equal-sized groups?

If so,  you can use this JSL* after you run kmeans for a one-cluster solution and save the distances:

((Col Rankc(:Distance Formula ) / Col Number( :Distance Formula )) * N Row() / :Group Size)

(*adapted from Phil Kay's post .)  

 

Re: Grouping by fixed number(i.e. 4 or 5) of products in cluster.

Due to the earnest request of the customer, I could not share the original data, so I created a similar data structure.

Please give your ideas

Thanks

Ikju

Re: Grouping by fixed number(i.e. 4 or 5) of products in cluster.

@jiancao Would this be another instance of the use of the covariate DoE, using blocks as mentioned in that thread about the mice? I haven't compared what is happening behind the scenes in covariate DoE with the clustering methods, but it seems the application is similar: optimally create equivalent groups of defined size.

Apologies, related thread here.  I've also attached the file used for exploring the idea. Only 1 covariate, weight, in this case (though the source table has 2 so i could explore different distributions).