Orthogonal Keys High Precision and Recall for Mining Database Keys From Inconsistent and Incomplete Relations

We introduce the notion of orthogonality between column sets as a means to eliminate accidental keys during data driven key discovery. This becomes particularly important for inconsistent relations, where keys that <i>almost</i> hold need to be considered as well. Here we employ the classic <inline-formula><tex-math notation="LaTeX">$g_{3}$</tex-math></inline-formula> metric for dirtiness, which can be applied to a variety of key semantics for incomplete relations, although the difficulty in computing the metric under different semantics varies greatly. Efficient algorithms for orthogonal mining and measuring dirtiness are proposed and evaluated on a real-world database. Additionally, we propose the notion of partial key, which turns out to be particularly useful for dealing with incomplete data in a key discovery context.