I have about 79,000 game players' data and we are trying to cluster these players into different classes. But so far we did not get a consistent cluster solution (we used K-means clustering). I plotted the data with two almost uncorrelated important variables identified by PCA, shown below. I also tried log transformation of the variables, which did not help much either. any suggestions? Maybe it is impossible to have a little overlapped clustering.
I used SAS to run the cluster analysis, and the ccc/pseudo F/Pseudo T^2 indicate there are 4 clusters:
but most of the players were not included in these clusters (because the "Distance Between Centroids" is 0 for them ):
So, does this mean there is no way to separate them into different groups?