I have a data set of size (4600, 10000). I did L2 normalization at first, then I did the following two steps to visualize it in a lower dimension:

1: Performed SVD and obtain 60 components, then performed PCA and obtain 30 components then did t-SNE on these 30 PCA components and obtain 2 components for visualization
2: Did PCA and obtain 50 components, then performed t-SNE and obtain 2 components for visualization.

The visualization from the first step is more clear than the second one. Does it make sense to do so? I all of these modules from scikit-learn.

  • 2
    $\begingroup$ You might be interested in learning more about the close relationship between PCA and SVD. stats.stackexchange.com/questions/134282/… $\endgroup$ – Sycorax Aug 26 '19 at 15:45
  • $\begingroup$ Yes, I tried to look at this post before as well, but still, I cannot decide either my step make sense or not. Thanks. $\endgroup$ – hemanta Aug 26 '19 at 15:48
  • $\begingroup$ The sense should come from the context of the problem you are trying to solve with your data reduction exercise. Which decomposition is more useful? In which decomposition there is more sense relative to the original features? $\endgroup$ – Tommaso Aug 26 '19 at 15:58
  • $\begingroup$ Actually, my goal is to do visualize such high dimensional data. So, I need to perform some dimensionality reduction techniques, for that, I choose PCA first but the components I obtained does not learn much. Then I tried the first step and is able to learn the data. I think doing SVD reduce the noise present in the data first? $\endgroup$ – hemanta Aug 26 '19 at 16:01

Your Answer

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

Browse other questions tagged or ask your own question.