Democratizing Self-Service Data Preparation through Example Guided Program Synthesis,

Abstract

The majority of real-world data we can access today have one thing in common: they are not immediately usable in their original state. Trapped in a swamp of data usability issues like non-standard data formats and heterogeneous data sources, most data analysts and machine learning practitioners have to burden themselves with "data janitor" work, writing ad-hoc Python, PERL or SQL scripts, which is tedious and inefficient. It is estimated that data scientists or analysts typically spend 80% of their time in preparing data, a significant amount of human effort that can be redirected to better goals. In this dissertation, we accomplish this task by harnessing knowledge such as examples and other useful hints from the end user. We develop program synthesis techniques guided by heuristics and machine learning, which effectively make data preparation less painful and more efficient to perform by data users, particularly those with little to no programming experience.

Data transformation, also called data wrangling or data munging, is an important task in data preparation, seeking to convert data from one format to a different (often more structured) format. Our system Foofah shows that allowing end users to describe their desired transformation, through providing small input-output transformation examples, can significantly reduce the overall user effort. The underlying program synthesizer can often succeed in finding meaningful data transformation programs within a reasonably short amount of time. Our second system, CLX, demonstrates that sometimes the user does not even need to provide complete input-output examples, but only label ones that are desirable if they exist in the original dataset. The system is still capable of suggesting reasonable and explainable transformation operations to fix the non-standard data format issue in a dataset full of heterogeneous data with varied formats. PRISM, our third system, targets a data preparation task of data integration, i.e., combining multiple relations to formulate a desired schema. PRISM allows the user to describe the target schema using not only high-resolution (precise) constraints of complete example data records in the target schema, but also (imprecise) constraints of varied resolutions, such as incomplete data record examples with missing values, value ranges, or multiple possible values in each element (cell), so as to require less familiarity of the database contents from the end user.