Hyperparameter tuning? Is this relevant in practice? Is it not rather an academic
gimmick? That the latter is not the case has been known for many years. On the other
hand, it is mostly unclear what exactly this looks like in practice. Which procedures
depend on which hyperparameters? How sensitive are the procedures to different
settings of their hyperparameters? And does that in turn depend on which data
constellations are available? How can users develop a good feeling for being on
the right track when tuning? Answers to these questions are not only expected when
it comes to optimally performing tuning per se, but also when it comes to making
the tuning process transparent, i.e., answering the question why, after all, this and
not that hyperparameter constellation was chosen.
This book delivers answers to the above questions, some of which were compiled
as part of a study funded by the Federal Statistical Office of Germany. The
contributed case studies and associated scripts also enable practitioners to reproduce
the described tuning procedures and apply them themselves. The presented
insights, cross-references, experiences, and recommendations will contribute to a
better understanding of hyperparameter tuning in machine learning and to gain
transparency.
