Experiment databases

Inductive databases are usually viewed as databases that contain models, next to data, such that mining the data essentially boils down to querying them for models. This unifies data mining with database querying. In his chapter, we consider experiment databases. Experiment databases are similar to regular inductive databases, but operate at a higher level of abstraction. They store not necessarily the models themselves, but properties of those models, and of the learners and datasets used to construct them. Querying them will reveal information about the behavior of data mining algorithms and the models they result in, when run on particular datasets. Thus, experiment databases unify meta-learning (learning about data mining systems) with database querying. In this chapter we discuss the advantages that experiment databases can offer in terms of openness of research, collaboration, reproducibility of experiments, thoroughness of analysis of experimental results, and correctness of their interpretation. We also describe the (principled) design of a pilot experiment database, and illustrate its use.status: publishe