Estimating the Quality of Data in Relational Databases

With more and more electronic information sources becoming widely available, the issue of the quality of these, often-competing, sources has become germane. We propose a standard for rating information sources with respect to their quality. An important consideration is that the quality of information sources often varies considerably when specific areas within these sources are considered. This implies that the assignment of a single rating of quality to an information source is usually unsatisfactory. Of course, to the user of an information source the overall quality of the source may not be as important as the quality of the specific information that this user is extracting from the source. Therefore, methods must be developed that will derive reliable estimates of the quality of the information provided to users, from the quality specifications that have been assigned to the sources. Our work here bears on all these concerns. We describe an approach that uses dual quality measures that gauge the distance of the information in a database from the truth. We then propose to combine manual verification with statistical methods to arrive at useful estimates of the quality of databases. We consider the variance in quality by isolating areas of databases that are homogeneous with respect to quality, and then estimating the quality of each separate area. These composite estimates may be regarded as quality specifications that will be affixed to each database. Finally, we show how to derive quality estimates for individual queries from such quality specifications. This work was supported in part by DARPA grants N0014-92-J-4038 and N0060-96-D-3202.