Bringing together visual analytics and probabilistic programming languages

Kurzfassung

A probabilistic programming language (PPL) provides methods to represent a probabilistic model by using the full power of a general purpose programming language. Thereby it is possible to specify complex models with a comparatively low amount of code. With Uber, Microsoft and DARPA focusing research efforts towards this area, PPLs are likely to play an important role in science and industry in the near future. However, in most cases, models built by PPLs lack appropriate ways to be properly visualized, although visualization is an important first step in detecting errors and assessing the overall fitness of a model. The thesis at hand aims to improve this situation by providing an interface between a popular PPL named PyMC3 and the software Lumen which provides several visualization methods for statistical models. The thesis shows how arbitrary models built in PyMC3 can be visualized with Lumen using the interface. It becomes clear that even for very simple cases, visualization can contribute an important part in understanding and validating the model since Bayesian models often behave unexpectedly. Lumen can therefore act as a useful tool for model checking.