Class imbalance in clustering Is there is a problem for clustering if the dataset is highly imbalanced? 
I have a clustering task and it looks like that there is a realy huge peak whose tail covers other clusters.
Are there any techniques to deal with that?
 A: In general: yes, this could very well be problematic. Imagine you have a number of clusters of unknown, but different classes. Clustering is usually done using a distance measure between samples. Many approaches thereby implicitly assume that the clusters share certain properties, at least within certain boundaries - like distances between clusters only diverging within a certain maximum, or more likely the scale of cluster spread only diverging within a certain maximum. This can be problematic in case you e.g. have one prominent, largely and unequally scattered class, that to some extent shadows other, less prominent, and as well unequally scattered classes - which further have very different distances between clusters. This could lead to your not-so-prominent clusters to not be found at all, as they e.g. get just pushed away from the prominent cluster (e.g. K-means), or could end up at just some slightly-above-average area of the prominent class (e.g. SOM, to some extent). But if this is the case with your problem, clustering will pretty likely be quite difficult with any clustering approach.
Two thoughts about possible approaches: 


*

*If you don't have any idea about the class prevalence, changing the data/data weight (e.g. subsampling using the density observed in your data) might defeat the purpose of clustering (imagine the extreme scenario of flattening out the whole feature space, which means discarding the information you would need for building clusters). But it could be that there are scenarios where this makes sense.

*If you have a rough idea of class prevalence, as @hxd1011 mentioned, using some weight for your clusters/distributions could be helpful. I guess that adapting the prevalence, using sampling techniques, the estimated prevalence, and the density observed in your data might be possible too (but keep in mind that when you use a mixed, observed density of different classes, your assumptions and simplifications might not be completely true, as mentioned in the first part of the answer). 
A: https://www.researchgate.net/post/Can_anyone_recommend_algorithms_to_deal_with_unbalanced_clusters_for_classification
I think "try agglomerative ones (single link, complete link, Ward, etc.)" may be a good answer
