FairPrep: fairness-aware human feedback for table unionability

The first adaptation of the FairEM framework from entity matching to annotator-level fairness in table union search, finding statistically significant bias against non-native English annotators on hard questions.

Human feedback supervises table union search in modern data lakes, but nobody had checked whether that feedback is equally reliable across annotator demographics.

Nine source modules across five experimental phases on a 480-row dataset; 10 FairEM confusion-matrix metrics, 10,000-shuffle permutation tests, 10,000-resample stratified bootstrap intervals, Fisher exact tests, Benjamini-Hochberg FDR correction and dual-track logistic regression.

1. 01

   6 of 10 FairEM metrics flag unfairness; 25% versus 44% accuracy on hard questions (p = 0.031).

2. 02

   Triple intersection of non-native, lower-education and young annotators persists at p_BH = 0.037 after FDR correction.

3. 03

   Findings robust across 5 of 5 orthogonal cleaning strategies; behavioral disparities exceed demographic ones.