It's not perfect, but you can put counts of points in each cluster (maybe you have count = 5, A_count = 2, B_count = 3 which means that this particular cluster has 5 points total, 2 of which are A and 3 of which there are B). Then you can filter based on these attributes.

The obvious downside of this is that the cluster positions aren't accurate because they're based on points that are supposed to be hidden, but it's fast and doesn't require recalculating.