Isotonic regression is useful when predictions need to respect an ordering. It finds values that are as close as possible to the original observations while remaining monotonic.
In production systems, it can be a practical way to repair a sequence of model outputs without hand-writing a large set of special-case rules.
