0
$\begingroup$

I am looking to employ an unsupervised clustering on a dataset where each observation has a mix of textual and non-textual features.

For each observation, I combine the features into a single vector of ~1000 dimensions. To cluster I have two potential ideas:

  1. Using an autoencoder (or an embedding?) to reduce the dimensionality of the data and then cluster using k-means.
  2. Could I use a topic model? If so, isn't this the superior method in most circumstances to the above?

Why are topic models (in my experience) not commonly used for non-textual data? Is this just a relic of their name/original application, or is there something more fundamental?

Thanks!

$\endgroup$
0
$\begingroup$

I think that you can use a topic model such as Latent Dirichlet Allocation (LDA). For example, in this paper https://pdfs.semanticscholar.org/9e6f/33bdd04df0536f6ad6783d33cccfbc54b1b1.pdf it is used for music and images. I suggest you to take a look at it :) . In general, in topic modeling you end up with a list of topics, where each topic contains a set of associated keywords. In clustering, depending on the algorithm, you might have hierarchy of dependencies. You can also use algorithms which assign each sample to only one class. In addition to this, when doing clustering, you usually have a distance metric which you have to pre-define (e.g. Euclidean distance). The topic models, especially LDA are based on the assumption that your data represents a distribution of topics with their corresponding distribution of keywords (one keyword can be contained in many topics). In other words, you already assume how the texts/documents have been generated.

$\endgroup$

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.