CrowdAidRepair: A crowd-aided interactive data repairing method

Abstract

Data repairing aims at discovering and correcting erroneous data in databases. Traditional methods relying on predefined quality rules to detect the conflict between data may fail to choose the right way to fix the detected conflict. Recent efforts turn to use the power of crowd in data repairing, but the crowd power has its own drawbacks such as high human intervention cost and inevitable low efficiency. In this paper, we propose a crowd-aided interactive data repairing method which takes the advantages of both rule-based method and crowd-based method. Particularly, we investigate the interaction between crowd-based repairing and rule-based repairing, and show that by doing crowd-based repairing to a small portion of values, we can greatly improve the repairing quality of the rule-based repairing method. Although we prove that the optimal interaction scheme using the least number of values for crowd-based repairing to maximize the imputation recall is not feasible to be achieved, still, our proposed solution identifies an efficient scheme through investigating the inconsistencies and the dependencies between values in the repairing process. Our empirical study on three data collections demonstrates the high repairing quality of CrowdAidRepair, as well as the efficiency of the generated interaction scheme over baselines.