Manually detecting errors for data cleaning using adaptive crowdsourcing strategies

Current work to detect data errors often uses (semi-)automatic solutions. In this paper, however, we argue that there are many real-world scenarios where users have to detect data errors completely manually, and that more attention should be devoted to this problem. We then study one instance of this problem in depth. Specifically, we focus on the problem of manually verifying the values of a target attribute, and shows that the current best solution in industry, which uses crowdsourcing, has significant limitations. We develop a new solution that addresses the above limitations. Our solution can find a much more accurate ranking of the data values in terms of their difficulties for crowdsourcing, can help domain experts debug this ranking, and can handle ambiguous values for which no golden answers exist. Importantly, our solution provides a unified framework that allows users to easily express and solve a broad range of optimization problems for crowdsourcing, to balance between cost and accuracy. Finally, we describe extensive experiments with three real-world data sets that demonstrate the utility and promise of our solution approach.