CoCo: Interactive Exploration of Conformance Constraints for Data Understanding and Data Cleaning
Data profiling refers to the task of extracting technical metadata or profiles and has numerous applications such as data understanding, validation, integration, and cleaning. While a number of data profiling primitives exist in the literature, most of them are limited to categorical attributes. A few techniques consider numerical attributes; but, they either focus on simple relationships involving a pair of attributes (e.g., correlations) or convert the continuous semantics of numerical attributes to a discrete semantics, which results in information loss. To capture more complex relationships involving the numerical attributes, we developed a new data-profiling primitive called conformance constraints, which can model linear arithmetic relationships involving multiple numerical attributes. We present CoCo, a system that allows interactive discovery and exploration of Conformance Constraints for understanding trends involving the numerical attributes of a dataset, with a particular focus on the application of data cleaning. Through a simple interface, CoCo enables the user to guide conformance constraint discovery according to their preferences. The user can examine to what extent a new, possibly dirty, dataset satisfies or violates the discovered conformance constraints. Further, CoCo provides useful suggestions for cleaning dirty data tuples, where the user can interactively alter cell values, and verify by checking change in conformance constraint violation due to the alteration. We demonstrate how CoCo can help in understanding trends in the data and assist the users in interactive data cleaning, using conformance constraints.
SIGMOD 2021 Demonstration Paper: CoCo: Interactive Exploration of Conformance Constraints for Data Understanding and Data Cleaning