How to accomplish unsupervised separation of subpopulations? I have a dataset drawn from a social network that looks Bimodal on logarithmic scales for all attributes (I'll demonstrate only one here):

I know the variable that would give me a clean separation for the two subpopulations (e.g., gender):

However, I need to come up with the same separation of the two clusters without using that binary variable (I could use internal knowledge of the two sub-populations statistics such as averages, stds etc .. ) in an unsupervised manner. 
I tried using Gaussian Mixture Models with EM, however, it didn't perform well on this dataset. Clearly what seems normal on logscale it isn't in reality :P 
I am looking for a package in python, R or Matlab that would be able to assign individual points to those two clusters. 
 A: If you think the attributes can reliably predict the two populations and you are trying to utilize all attributes to achieve unsupervised separation, then you are describing a typical clustering problem.  It is fairly simple to get started in Python with scikit-learn. 
If you have a gold standard of labeled data (vectors of attribute values that characterize each of the two populations), you can also do this in a supervised manner with any of the many classification algorithms (also available in scikit-learn).  Depending on the dimensionality, RandomForestClassifier or support vector machine methods like LinearSVC are good starting points. 
A: I'd strongly recommend you google semi-supervised learning. It is directly designed to deal with situations where there is a small labelled set and a much larger unlabelled set, and you try to learn from both. In this situation, I feel that training your model with all the data but no labels (unsupervised) or with labels but not all the data (supervised) is very wasteful. Why not use everything you have?
I'm not very knowledgeable about the details of semi-supervised algorithms so I think you are probably better off spending a bit of time reading about it from review articles you can find online than listening to me. This paper is a little old by ML standards (2007) but it's very heavily cited and seems solid: http://aims.cse.msu.edu/~cse802/notes/SSLsurvey.pdf
