Сlassifying clusters

Question

Let's suppose we have something like a classification problem. We have a training dataset where every record has its class label. There are several classes and classes don't overlap. The data that model should be making predictions for will be divided into clusters and we know that every cluster corresponds to a single class label from a training set (all points in one cluster are drawn from the same class, but we don't know which one).

Is there a specific name for such kind of problems, where we don't actually have to predict a class label for every record but rather predict class label for a whole cluster? If no, what approaches are best suited for this task?

Of course, we can treat this as a common classification problem just by predicting a class label for every record in a cluster and than choosing the most popular prediction as a prediction for a cluster. But I feel like this approach does not take into account general distribution of features in each cluster so it does not use all the information that is avaliable for prediction

Answer 1

The great benefit of unsupervised learning (e.g. clustering) is that you do not need labeled data. Therefore, just from the features of each data point and the distances between them, you can generate groups/clusters with smaller or greater degree of separability. For this reason, I don't see the point of doing clustering on labeled data. It defeats the purpose of the technique.

Maybe I am missing some detail here, and someone else can bring an interesting input, but: why would you want to generate clusters (which number by the way is generally unknown) on labeled data?