Isotonic regression builds an increasing approximation of a function while maximizing the mean squared error on training data. In the case of linear regression, it minimizes the mean squared error. This kind of model is useful when you do not make any kind of assumption for target function such as linearity.