I have large data (more 200K with more 30 feature). I try clusteirng model as K-means, Gaussian Mixture, Agglomerative hierarchical clustering. Overall, all give same result. I use under 25K datasets that still provide varied clusters but when more than 30K datasets only 99% of data is in one cluster. The number of cluster is 6, it the best value from Elbow method.
Hi Yusuf,
I don`t know what software you are using, but if you are using R you can control for your clusters to have approximately equal size following this:
https://www.r-bloggers.com/equal-size-kmeans/
I believe this can be done in other statistical softwares.
Just a comment, if k-means is classifying 99% of your data into one group this must mean something. Perhaps, the results your getting are not a bad thing and you can take more interesting insights from that than having your data classified into similar size clusters.
Hope it helps!
1-Try elbow plot
2-Take a close look at the dendrogram, check the linkage height
Thank you all for it. I had solved the problem with dummies approach on numerical dataset, So I create new categorial data from numerical data. Moreover, I used principal component analysis (PCA) for extraction feature. The cluster provide good results with largest class member 20% of the data.
- Badges
- Science topic