Deriving Effectiveness Measures for Data Quality Rules

The poor quality of data constitutes a major concern worldwide, and an obstacle to data integration and analysis efforts. Detecting errors and inconsistencies using application specific data quality rules play an important role in data quality assessment. These rules have different efficacy and cost under different circumstances. In our previous work, we have proposed a quantitative framework for measuring and comparing data quality rules in terms of their effectiveness. Effectiveness formulas are built from variables that represent probabilistic assumptions about the occurrence of errors in data values, and our earlier work gave examples of how to derive these formulas in an ad-hoc fashion. This paper lays the foundations of a workbench-approach for systematically deriving effectiveness formulas. The approach involves several steps, including building Bayesian network graphs, adding (symbolic) probabilities to the nodes in the graph, and deriving effectiveness formulas. The graphs are built algorithmically, for a large and useful class of data quality rules. We present this approach and its implementation in Python, and report its evaluation results, which show that the resulting formulas give reasonable estimates of effectiveness scores under various scenarios.