Generating the clustering results using Minisom about minisom

[JustGlowing]

hi @soumya-ranjan-sahoo, this question has been already answered here: #25

[soumya-ranjan-sahoo]

Hi Team,

Thanks for a prompt response. I might need some more clarity on this one - For instance, if I want to cluster my data points into n-clusters, I need to design the SOM with n1 nodes on the output map? How about having an nm map where data points belonging to the same cluster are spread across the grid (Eg. in Iris data set, let's say orange, green and blue are 3 clusters and they are spread across the grid i.e. multiple nodes have orange clusters. How do we go about this?)

[JustGlowing]

Then you have to treat the coordinates on the map as features and use another clustering algorithm to get the clusters.

[soumya-ranjan-sahoo]

If I treat the coordinates as features, then the grid-based distance between the nodes might be the same for positive as well as the negative class datapoints. For instance, let's have 3 nodes (0,0),(0,1),(0,2) with 2 classes. Class 1 might be mapped to (0,0) and (0,1), while class 2 gets mapped to (0,2). In this case, taking (i,j) coordinates as features how do I differentiate between 2 class as the distance between (0,0) and (0,1) is the same as (0,1) and (0,2).

Also, if I want to cluster my data points into n-clusters, do you recommend using a 1-D som of shape (1,n) where n = number of clusters?

Thanks in advance!

[JustGlowing]

Hi again @soumya-ranjan-sahoo,

In the case that you described, where two the classes are equidistant, you should pick one at random unless you have some prior knowledge on the problem.

About the map size:

• If you want to cluster on the space created by the som, I'd recommend you to create a square map. Possibly large enough to map a sample per cell.
• If you want to use a small map, just consider each neuron as a cluster (as suggested in #25)

[JustGlowing]

hi there, I made this example to show how to use minisom for clustering:

https://github.com/JustGlowing/minisom/blob/master/examples/Clustering.ipynb