Top-k join queries: overcoming the curse of anti-correlation

The existing heuristics for top-k join queries, aiming to minimize the scan-depth, rely heavily on scores and correlation of scores. It is known that for uniformly random scores between two relations of length n, scan-depth of √kn is required. Moreover, optimizing multiple criteria of selections that are anti-correlated may require scan-depth up to (n + k)/2. We build a linear space index, which in anticipation of worst-case queries maintains a subset of answers. Based on this, we achieve Õ(√kn) join trials i.e., average case performance even for the worst-case queries. The experimental evaluation shows superior performance against the well-known Rank-Join algorithm.