Machine learning methods
for global model induction, like support vector machines or artificial neural
networks, are nowadays applied in a wide range of data-driven applications.
Therefore they appear like a natural tool also for scientific data analysis.
However, although their models can reach astounding accuracies, they tend to
offer surprisingly little insight into the underlying domain.
Local modeling methods address this concern by being potentially agnostic about
parts of the input space in order to focus on specific effects that can be
modeled in simple terms with high precision---in particular those that are
unusual or outstanding given the global picture. Additionally, they achieve
interpretability by using discrete symbols that correspond to meaningful
notions of the discovery domain.
In this talk, I present a representative local modeling technique called
subgroup discovery. I show how it was successfully used in scientific
applications and discuss its computational complexity as well as practically
effective algorithms.
