摘要: Data analysts often need to transform an existing dataset, such as with filtering, into a new dataset for downstream analysis. Even the. most trivial of mistakes in this phase can introduce bias and lead to the formation of invali... 展开 Data analysts often need to transform an existing dataset, such as with filtering, into a new dataset for downstream analysis. Even the. most trivial of mistakes in this phase can introduce bias and lead to the formation of invalid conclusions. For example, consider a researcher identifying subjects for trials of a new statin drug. She might identify patients with a high dietary cholesterol intake as a population likely to benefit from the drug, however, selection of these individuals could bias the test population to those with a generally unhealthy lifestyle, thereby compromising the analysis. Reducing the potential for bias in the dataset transformation process can minimize the need to later engage in the tedious, time-consuming process of trying to eliminate bias while preserving the target dataset. We propose a novel interaction model for explain-and-repair data transformation systems, in which users interactively define constraints for transformation code and the resultant data. The system satisfies these constraints as far as possible, and provides an explanation for any problems encountered. We present an algorithm that yields filter-based transformation code satisfying user constraints. We implemented and evaluated a prototype of this architecture, EMERIL. using both synthetic and real-world datasets. Our approach finds solutions 34% more often and 77% more quickly than the previous state-of-the-art solution. 收起