I am trying to simultaneously cluster and visualize text documents using Self-organizing maps. Since text documents can be represented in various ways (vector space model, GloVe etc), I am trying to figure out how to tell which representation generates the best map. Measures like Quantization error etc., determine the goodness of the map given a dataset. However, they are not useful for quantitatively telling which representation gives a better output.

Is there a quantitative measure to compare the maps generated using different representations (for example, Tf-idf and GloVe) and tell for which representation the output is better?

  • $\begingroup$ Use something else, SOM are so outdated... A basic autoencoder does better and has a more continuous nD space. $\endgroup$ – Matthieu Brucher Jan 22 '19 at 14:16

From Wikipedia:

A self-organizing map (SOM) or self-organizing feature map (SOFM) is a type of artificial neural network (ANN) that is trained using unsupervised learning to produce a low-dimensional (typically two-dimensional), discretized representation of the input space of the training samples, called a map, and is therefore a method to do dimensionality reduction.

So you only have the original data itself; no additional data (like labels in a supervised setting). If you are also say the result has to have two dimensions, you basically look at functions

$$f: X \rightarrow \mathbb{R}^2$$

where $X \subsetneq \mathbb{R}^n$ in most cases. You already mentioned quantization error.

Up to my knowledge there is nothing better measure which does not include getting more knowledge about the data itself by human inspection / using other datasets.

With human inspection you can, of course, tell for a given dataset and a given human if one mapping seems to make more sense.

You might also consider other dimensionality reduction techniques:


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.