Using Self Organizing Maps algorithm to cluster some data will give us NXM centroids where N and M are pre-defined map dimensions. If I have a distribution of species and want for example to get 4 centroids. I am not sure it is correct to choose for example 2x2 SOM dimensions.
Thus, I put 10x10 map then use the K-means to get 4 centroids. However, if I use 8x8 map I will get different centers. So, my question is what is the best NxM dimension?
Recently I had the same issue. However 1D SOM work well in my case... you can try it too.. i.e. 1xN
I utilized it in the following research
https://www.mdpi.com/1424-8220/20/3/873
.
i would not recommend constraining the map size by some cluster number you would request
som is an excellent tool for exploratory data analysis : you can learn much more than you would expect by examining the results of a large map (U-matrices will give you a lot of information)
then, it is always possible to cluster the neurons of map in order to obtain the number of groups you want ; this post-treatment can be made with hierarchical agglomerative clustering, k-means or any other clustering technique, it can take the population of each neuron into account or not
usually i use HAC and do not take the populations into account : it gives me groups of neurons with close features and the population gathered by each such group is a cluster in my population
.
Thank you @Amin Ullah, I am not sure if will work, my data are scattered points based on the longitudes and latitudes. I will try though.
Hi everyone,
I choose dimensions like this:
If I have a dataset with 30 attributes, I assign the number of rows = 5, number of cols = 6, dim = 30.
I try to create a grid as close to square as possible.
What do you think? Can this approach work?
Regards,
- Badges
- Science topic
- Similar topics
- Geoscience
- Cartography
- Mapping
- Maps