kmeans appear to miss obvious clusters

6 visualizaciones (últimos 30 días)
NCramer
NCramer el 7 de Jul. de 2017
Comentada: NCramer el 10 de Jul. de 2017
I am struggling to get kmeans to identify what appear to be fairly distinct clusters in my data. I've walked through the documentation and examples but can't improve over the images shown below (raw data plotted first followed by the kmeans result, data also attached). I've tried the different distance and start options without much success. Even giving seed values doesn't improve the clustering. Does anyone have any other suggestions I could try? My goal is to end up with each data point falling into one of 3 clusters. My last command was:
[cidx3,cmeans2] = kmeans(X,3,'dist','cosine','display','iter','Start',seeds);
where
seeds =
[0.018660 872 17.59;
0.002100 1140 18.88;
0.004652 1187 34.82]
Thank you
  1 comentario
Adam
Adam el 7 de Jul. de 2017
It would help if you plotted the seeds visibly on the graph. It's not very easy to see where a point is in 3d just from coordinates.

Iniciar sesión para comentar.

Respuesta aceptada

Ilya
Ilya el 7 de Jul. de 2017
Do this (assuming there are no nan's in X):
[cidx3,cmeans2] = kmeans(zscore(X),3,'dist','cosine','display','iter');
Did it get better? If yes, look at your data again and think about what went wrong in your previous attempts. Look at the scales. Plot it using real scales 1:1. Think about how the cosine distance works when the data are shifted far away from zero.
  1 comentario
NCramer
NCramer el 10 de Jul. de 2017
Using zscore to normalize the data did help significantly. Thank you (and Image Analyst) very much for that suggestion. There is still a fair amount of bleeding of the main cluster into the smaller ones but I will play around with other ways to normalize the data and see if that helps.

Iniciar sesión para comentar.

Más respuestas (1)

Image Analyst
Image Analyst el 7 de Jul. de 2017
You might want to normalize your data.
I don't think it's good to try to find clusters when one parameter goes from 0 to 1500 and another goes from 0 to 0.05 !!!
With these ranges, your data is basically in a skinny flat sheet, not a 3-D widely spread out space.
  1 comentario
NCramer
NCramer el 10 de Jul. de 2017
Thank you, Image Analyst. Normalizing with zscore does certainly help.

Iniciar sesión para comentar.

Etiquetas

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by