Can you explain how to group users even if they have very different statistical distribution?
It seems to be an unsupervised problem. I know that K-means is commonly used to do clustering but is K-means really effective when users have very different statistical distribution?
Thank you.
K-means indeed works best when all clusters have the same error distribution around their mean, i.e., if they all follow N(mu_j, eps).
But there are at least 100 clustering algorithms. Just pick one the suits your data better. For example DBSCAN only assumes that each cluster is dense, and clusters are separated by gaps of low density. It can't cluster overlapping Gaussians well, but it will work fine with well separated clusters of different distributions.