I need to know if the following problem was already adressed anywhere (and if so, then how is it called in literature).
Lets say, that we do have clustered datapoints (based on some distance function, that is not ideal) and there is a sequence of new datapoints incoming and we need to dynamically assign them to existing clusters or create new clusters. (I did try googling dynamic clustering, but that is not it).
There is, of course a training dataset, that would tell you if assigning a point to a (specific) cluster was bad and if creating a new cluster was bad (and it can measure the badness). The original distance function is some approximation to the training dataset, but can be evaluated anywhere. The training dataset is the main truth, but can be provided just for some data.
If there was never such a problem solved, I will have a nice topic for a scientific article :)