As part of a series of research projects aimed at marrying machine learning technology to economically important problems, this project in concerned with the task of finding errors in databases. Because many database situations do not afford the luxury of a surplus programmer, machine learning methods are used to construct prescriptive data models. These models predict appropriate values for data attributes, and if actual values differ, an alarm is raised. Two basic research issues in database consistency checking are studied. First, fundamental limitation of a common class of learning methods is identified, then a new learning formulation is proposed. Second, this research reiterates the need to incorporate expert knowledge into the learning process, and a space of possible approaches. Understanding of the resulting new methods will lead to advances in the quality and applicability of database consistency systems, and as a consequence, to improved performance in the wide variety of systems that rely on high quality data.
