1
$\begingroup$

I have a social science background and I'm doing a text mining project. I'm looking for advice about the choice of the number of topics/clusters when analyzing textual data. In particular, I'm analyzing a dataset of more than 200000 tweets and I'm performing a Latent Dirichlet allocation model on them to find clusters that represent the main topics of the tweets of my dataset. However, I was trying to decide the optimal number of clusters but the results I'm finding in the picture seem inconsistent. enter image description here

I'm struggling with the choice of the number of clusters. So the question is: what number would you choose from the plot? Moreover, do you think there are other ways and/or conventional rules that one can rely on to choose the number of clusters?

$\endgroup$
1
  • 1
    $\begingroup$ Welcome to DataScienceSE. There's no "right answer" about the number of topics in general. Quite often people manually check the top words for the different topics to decide whether the set of topics makes sense for the application. See also this related question. $\endgroup$
    – Erwan
    Sep 5, 2021 at 16:38

1 Answer 1

0
$\begingroup$

You can try Perplexity or Coherence Score for the number of topics.

  1. Perplexity is a statistical measure of how well a probability model predicts a sample. As applied to LDA, for a given value of k, you estimate the LDA model. Then given the theoretical word distributions represented by the topics, compare that to the actual topic mixtures, or distribution of words in your documents. - Lower the better.
  2. Coherence Score is defined as the average/median of the pairwise word-similarity scores of the words in the topic - The higher the better.
$\endgroup$

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.