Statistical Distortion: Consequences of Data Cleaning

Abstract

We introduce the notion of statistical distortion as an essential metric for measuring the effectiveness of data cleaning strategies. We use this metric to propose a widely applicable yet scalable experimental framework for evaluating data cleaning strategies along three dimensions: glitch improvement, statistical distortion and cost-related criteria. Existing metrics focus on glitch improvement and cost, but not on the statistical impact of data cleaning strategies. We illustrate our framework on real world data, with a comprehensive suite of experiments and analyses.