Towards Declarative, Domain-Oriented Data Analysis

The need for advanced data analysis now pervades all areas of science, industry and services. A wide variety of theory and techniques from statistics, data mining, and machine learning is available. Addressing a concrete question or problem in a particular application domain requires multiple non-trivial steps: translating the question to a data analysis problem, selecting a suitable approach to solve this problem, correctly applying that approach, and correctly interpreting the results. In this process, specialist knowledge on data analysis needs to be combined with domain expertise. As data analysis becomes ever more advanced, this becomes increasingly difficult. In an ideal world, data analysis would be declarative and domain-oriented: the user should be able to state the question, rather than describing a solution procedure, and the software should decide how to provide an answer. The user then no longer needs to be, or hire, a specialist in data analysis for every step of the knowledge discovery process. This would make data analysis easier, more efficient, and less error-prone. In this talk, I will discuss contemporary research that is bringing the state of the art in data analysis closer to that long-term goal. This includes research on inductive databases, constraint-based data mining, probabilistic-logical modeling, and declarative experimentation.