Towards automatically linking data elements

Abstract

When presented with a new dataset, human data scientists explore it in order to identify salient properties of the data elements, identify relationships between entities, and write processing software that makes use of those relationships accordingly. While there has been progress made on automatically processing the data to generate features or models, most automation systems rely on receiving a data model that has all the meta information about the data, including salient properties and relationships. In this thesis, we present a first version of our system, called ADEL-Automatic Data Elements Linking. Given a collection of files, this system generates a relational data schema and identifies other salient properties. It detects the type of each data field, which describes not only the programmatic data type but also the context in which the data originated, through a method called Type Detection. For each file, it identifies the field that uniquely describes each row in it, also known as a Primary Key. Then, it discovers relationships between different data entities with Relationship Discovery, and discovers any implicit constraints in the data through Hard Constraint Discovery. We posit two out of these four problems as learning problems. To evaluate our algorithms, we compare the results of each to a set of manual annotations. For Type Detection, we saw a max error of 7%, with an average error of 2.2% across all datasets. For Primary Key Detection, we classified all existing primary keys correctly, and had one false positive across five datasets. For Relationship Discovery, we saw an average error of 5.6%. (Our results are limited by the small number of manual annotations we currently possess.) We then feed the output of our system into existing semi-automated data science software systems - the Deep Feature Synthesis (DFS) algorithm, which generates features for predictive models, and the Synthetic Data Vault (SDV), which generates a hierarchical graphical model. When ADEL's data annotations are fed into DFS, it produces similar or higher predictive accuracy in 3/4 problems, and when they are provided to SDV, it is able to generate synthetic data with no constraint violations.