Smart Data Discovery
Concerning “Smart Data Discovery”, we are conducting research together with our scientific partners (e.g. RU Nijmegen) on the following topics, among others:
Constraint-based algorithms for causal inference are ideally suited for parallel processing. They prevent the search over all model structures by taking a stepwise, modular approach. Essentially, constraint-based algorithms test for conditional dependencies and independencies involving subsets of variables and translate the outcomes of these tests to causal statements. Each computation follows the same pattern at run-time, a crucial precondition for effective parallelization on, for example, GPUs.
To this end, we combine and compare causal discovery methods based on conditional independencies with other approaches that attempt to quantify the directed influence between concurrent processes. In particular, we study transfer entropy, an information-theoretic measure of information transfer. Such network models can then be used to spot structural design improvement areas and, in case of known failures, to identify potential root causes.
Trustworthy, Interpretable Models
With respect to explain individual outcomes of a learned model our research builds on methods like LIME which allow computing local explanations. Local explanatory models can be built by methods like rule induction and/or decision tree learning which can leverage expert knowledge formalized in ontologies. Hence the resulting local explanation model will be capable to “talk the language” of the domain experts.
Integration of Event / Log Data
A process execution model can be reproduced from event data using process mining. The findings may reveal e.g. bottlenecks in throughput times or manual reworking. Such identified problems may be considered as the new target to be explained. Therefore, we exploit the methods developed for causal discovery to find the root causes of such problems.
Selected publications are summarized in the strategic project dasRES.