When building classification models of complex systems with many classes, the traditional chemometric approaches such as discriminant analysis or soft independent modeling of class analogy often fail. Some people resort to advanced deep neural network, but this is only an option if there is access to very many samples. Another alternative often used is to build hierarchical models where subclasses are sort of peeled off one or a few at a time. Such approaches often outperform classical classification as well as deep neural network on small multi-class problems. The downside though is that it is very cumbersome to build such hierarchies of models. It requires substantial work of a skilled person. In this paper, we develop a fully automated approach for building hierarchical models and test the performance on a number of classification problems.