I am clustering tweets which are related to eye fashion and they are extracted using keywords like mascara, eyeliner, eyeshadow, etc from twitter. I constructed a Tf-idf matrix (tweets x words) which was around (100000 x 550).
If I give this to K-Means and plot elbow plot I don't get a perfect elbow. (Its almost diagonal) But after applying truncated SVD with latent variables say 10 variables and then I do kmeans, I get a perfect elbow at 10.
I have tested this for various other numbers of latent variables too and I get perfect elbow at those numbers. Can the results be trusted in this case?
I don't see the clusters having similar topics. What is the best way to proceed in this case. I did try LDA but there also I dont see perfect distinction between topics. Is it because that the data itself is not clusterable?