If we get rid of the label in the Iris dataset, it would be nice if some algorithm could recover the original grouping, maybe without the exact label names—setosa, versicolor, and virginica. Unsupervised learning has multiple applications in compression and encoding, CRM, recommendation engines, and security to uncover internal structure without actually having the exact labels. The labels sometimes can be given base on the singularity in attribute value distributions. For example, Iris setosa can be described as a Flower with Small Leaves.
While a supervised learning problem can always be cast as unsupervised by disregarding the label, the reverse is also true. A clustering algorithm can be cast as a density-estimation problem by assigning label 1 to all vectors and generating random vectors with label 0 (The Elements of Statistical Learning by Trevor Hastie, Robert Tibshirani, Jerome Friedman, Springer Series in Statistics). The difference between the two is formal...