How do I get the cluster number to which each sample belongs?

[sumanth-bmsce]

Hi,

Is there any way I can get the cluster number to which each datapoint in that dataset belongs to?

[JustGlowing]

Hi, since this is a recurring question I decided to add an example here: https://github.com/JustGlowing/minisom/blob/master/examples/Clustering.ipynb

[sumanth-bmsce]

Hi,

Thanks for quickly answering my question. Depending on the dataset info i.e. no of clusters and number of data points in each class, we can calculate the centroids for the coordinates returned by som.winner and use it to find the cluster number and calculate the clustering accuracy?

[JustGlowing]

Of course you can. The centroids are the weights of the som (you can also compute the centroids by averaging the data in each cluster). You can use any metric for computing the accuracy of your clustering method, minisom offers the quantization error.

…

On Sat, Mar 28, 2020, 10:38 Sumanth Simha C ***@***.***> wrote: Hi, Thanks for quickly answering my question. Depending on the dataset info i.e. no of clusters and number of data points in each class, we can calculate the centroids for the coordinates returned by som.winner and use it to find the cluster number and calculate the clustering accuracy? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#64 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFTNGISOJJXO5RAMEOJEUDRJXHSHANCNFSM4LVQR5BA> .

[sumanth-bmsce]

Is it possible that I can contribute computing the cluster numbers and clustering accuracy in the way which I described before to this repo?

[JustGlowing]

Of course, you have to open a pull request. Google is your friends for that ;)

…

On Sat, Mar 28, 2020, 11:06 Sumanth Simha C ***@***.***> wrote: Is it possible that I can contribute computing the cluster numbers and clustering accuracy in the way which I described before to this repo? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#64 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFTNGLW7CQQDV7OQXOVCVLRJXK23ANCNFSM4LVQR5BA> .

[JustGlowing]

PS: I updated the example so that also the centroids are plotted.

[Edmond-Lee-Zse-Wong]

Hi,
my original dataset has no labels, is there any way to sign every datapoints after som training?
PS: I use a 9*9 som and the distance map shows that the dataset is approximately divided to 2 clusters
Thx !

[JustGlowing]

hi @Edmond-Lee-Zse-Wong check the example linked above.

[Edmond-Lee-Zse-Wong]

Hi
I've checked it last night but the problem hasn't been tackled. As you can see 11 winner neurons are activated, but my dataset has no possibility to be divided to 11 clusters since it is a simple bidimensional dataset(236 points). The som size is calculated by an experiance formulation: size = 5√N, where N is the number of datapoints. The only solution is to reduce the size of som?

[JustGlowing]

It's likely that some clusters are spread across different neurons. The som size is not adequate for the task that you are trying to perform.

[Edmond-Lee-Zse-Wong]

Yes I think so, maybe I should reduce the size to 2 or 3.
Thx for your help !!!