Measuring approximate functional dependencies: a comparative study

Abstract

Approximate functional dependencies (abbreviated: AFDs) are functional dependencies (FDs) that "almost" hold in a relation. While various measures have been proposed to quantify the level to which an FD holds approximately, they are difficult to compare and it is unclear which measure is preferable when one needs to discover FDs in real-world data, i.e., data that only approximately satisfies the FD. In response, this paper formally and qualitatively compares AFD measures. We obtain a formal comparison through a novel presentation of measures in terms of Shannon and logical entropy. Qualitatively, we perform a sensitivity analysis w.r.t. structural properties of input relations. Quantitatively, we study the effectiveness of AFD measures for ranking linear AFDs on real world data. Based on this analysis, we give clear recommendations for the AFD measures to use in practice.