Compiling Relational Database Schemata into Probabilistic Graphical Models

A majority of scientific and commercial data is stored in relational databases. Probabilistic models over such datasets would allow probabilistic queries, error checking, and inference of missing values, but to this day machine learning expertise is required to construct accurate models. Fortunately, current probabilistic programming tools ease the task of constructing such models [1, 2, 3, 4, 5, 6] and work in statistical relational learning has focused on making it even easier to define models specific to relational data [7, 8, 9, 10]. However, within these frameworks the user still needs to specify all the probabilistic dependencies in the data, requiring a level of expertise in probability and statistics that domain experts often do not have, thus severely restricting the practical applications of such techniques. On the other hand, domain experts do spend considerable effort and expertise in designing the database schemata used to represent their data, providing type information for table columns and foreign key relations to specify dependencies.