How do I create balanced clusters?

kikikubikova

Hi guys,

I'm pretty new to the community so sorry if my question will seem quite elementary, but how do I create balanced clusters (k-means) - meaning that each cluster will have the same size of items in it? Or is there a way to force a minimum cluster size to anything else than 1? 

(What I am trying to do is to create pairs based on some variables - I have a list of villages, their population size, average age, unemployment etc. And for each village in my dataset I am looking for the village with the most similar parameters in all of the variables - matching the most alike villages. My idea was to do N/2 clusters to create pairs, but as I don't know how to do balanced clusters or how to force the minimum size of a cluster to 2 items, the output was N/2 clusters but unfortunatelly there weren't 2 items in each, creating some clusters with i.e. 3 items and some with 1 item in it.)

Thank you for all of your advices (the simpler solution the better ) !

MartinLiebig

Hi,

let me ask the question first: Why do you want to cluster? What do you want to do with the similar villages?

BR,

Martin

kikikubikova

Hi, 

thank you for your comment. So, as my master thesis I am analyzing the effect of rainfall on voter turnout. I have turnout and precipitation data for around 600 villages as well as some basic information like unemployment, area, population size etc. My task is to match the villages based on the parameters, finding the most suitable pair to perform a diff-in-diff model (checking how the difference in turnout changes with the difference in precipitation and other independent variables between the two villages during years).

Do you have any idea how to fix the number of items in a cluster? Or how to increase the minimum number of items in a cluster to 2?

Thanks,

Kristina

MartinLiebig

Hi,

so why do you think clustering would be a good approach here?

I would either:

- Build a predictive model to predict turn out and then look at the influence of rain fall, for example with PDP plots.

- Search for the k-Next neighbours of each village and then compare their turnout with the turnout in this one village.

For normal K-Means you are not able to control cluster sizes, other than merging "close" clusters until they got big enough. That would be a manual process I think. Anyway, I do not think clustering is a suitable approach here.

BR,

Martin

kikikubikova

With all due respect, Martin, that was not my question. I wasn't asking about your opinion on my model and methodology, I was asking how to do balanced clusters using RapidMiner. For your information, you are able to control cluster sizes using k-means (i.e. DOI:10.1007/978-3-662-44415-3_4 or simply Wikipedia "Balanced clustering"), and you can do so using for example R or Python. My question was whether it is possible, and if so how to do it, using RapidMiner. 

Anyway, thank you for your adivce!

BR,

Kristina

MartinLiebig

Then: Balanced K-Means is not available in RM natively.