I'm currently working on a research of data clustering using an ANN for self-organizing maps. I'm performing experiments using Matlab, over a Dataset of 20,000 samples and almost 80 variables.

The ANN requieres to set up a "one or two dimensional lattice", which is used to place neurons at each node. this map's function is to transform an arbitrary signal (with n dimensions) into a discrete representation with topological order.

I haven't found yet a rational criteria to select a lattice with one or two dimensions, it's worth noting that a node does not represent necessarilly a cluster because you can group a set of nodes in one cluster (it depends on how close are each other).

Does anybody knows a good criteria to set up the lattice? Greetings!

  • $\begingroup$ This might come down to interpretation. Does it make sense to cluster solely on a single axis? Think: hot/cold, conservative/liberal, high/low. If that seems restrictive to you then maybe a single axis might not be enough (not to mention it would be somewhat unconventional) $\endgroup$ – shadowtalker Feb 24 '15 at 2:56
  • $\begingroup$ First of all, let offer you some apologies, my question wasn't clear at all, I've already change it, hope it fits better. Actually, in a SOM clustering process, the lattice's dimensions doesn't represent a single variable, on the contrary, the lattice is a representation of n variables which conserves the original topological order of the data. Thank you so much $\endgroup$ – formacero10 Feb 27 '15 at 1:00
  • $\begingroup$ exactly. Hence, my question: does it make substantive sense to use a 1-D lattice? $\endgroup$ – shadowtalker Feb 27 '15 at 1:14
  • $\begingroup$ I've concluded (sort of... I'm not sure) that it makes sense if you want a fixed ammount of clusters, so you simply define a lattice having as much nodes as clusters you need, but you lose quiality in the topological representation of the data. This quality lost is a clear argument to ditch a 1D lattice? $\endgroup$ – formacero10 Feb 27 '15 at 1:24

In my expirience there isnt an objective rule for the selection of the lattice. A 2d lattice results to better topological adaptation than an 1d lattice but then a 3d lattice gets even beter results. Since one of the advantages of the som map is the visual exploration of the results you go by that. A 3d lattice becomes kind of hard to explore so usually the standard choice is a 2d lattice. As ssdecontrol mentioned an 1ď lattice will give you a very "naive" relationship between the clusters - nodes.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.