Data Cleaning for Decision Support

Data cleaning may involve the acquisition, at some effort or expense, of high-quality data. Such data can serve not only to correct individual errors, but also to improve the reliability model for data sources. However, there has been little research into this latter role for acquired data. In this short paper we define a new data cleaning model that allows a user to estimate the value of further data acquisition in the face of specific business decisions. As data is acquired, the reliability model of sources is updated using Bayesian techniques, this aiding the user in both developing reasonable probability models for uncertain data and in improving the quality of that data. Although we do not deal here with the problem of finding optimal methods for utilizing external data sources, we do show how our formalization reduces cleaning to a well-studied optimalization problem.