I have a vector-space-model. I want to cluster the embeddings (based on similarity) in this model, but I have following dilemma / confusion: Should I

  1. first cluster in the higher dimensional space and then do dimensionality reduction to get a 2D plot I can display?


  2. first do the dimensionality reduction to a 2D space and then cluster on the 2D representations?


1 Answer 1


I would say, (1) is not an option. Vector-space-model usually result in hundreds of features. Most clustering algorithms rely on the concept of distance between data points. With hundreds of features, you suffer from the curse of dimensionality. This makes most distance functions less effective (to say it politely).

(2) is also not so good, because reducing from hundreds of features to 2 will definitely introduce large distortions.

Thus, I would say, first try to reduce dimensionality to something reasonable (depending on number of data points you have). Then cluster. Then think about possible ways to visualize in 2D.

  • $\begingroup$ I dont think curse of dimensionality is problem as essentially I have vector-space, which I have the basis for $\endgroup$
    – SFD
    Mar 6, 2018 at 22:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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