اكتب مراجعة على PREreview

Predictive models are now embedded across education, flagging at-risk students and shaping the support they receive. Most fairness audits measure group disparity in the final prediction and then apply a fix, but this end-of-pipeline view cannot say where the unfairness originated or how far it travelled. We introduce Bias Provenance and Propagation (BPP), a measurement framework—not a new mitigation method—that treats disparity as a quantity with a source stage and a path through the machine-learning pipeline. Five stages (sampling, preprocessing, feature engineering, model training, evaluation) plus an exploratory simulation of distribution shift are placed on one threshold-independent scale, the Absolute Between-ROC Area (ABROCA), by projecting each stage's information through a single frozen reference model. Two derived measures form a bias ledger that records where fairness debt is incurred and repaid, each with a bootstrap confidence interval. Applied to the Open University Learning Analytics Dataset (32,593 records) on a leakage-free early-warning task, BPP shows that the origin of disparity is attribute-specific: for disability it is present in the raw data and attenuates once early assessment and registration features enter, whereas for sex/gender and deprivation it is introduced at feature engineering. Several pre-specified hypotheses were not supported—in particular, no universal origin stage exists (Kendall's W = 0.11)—and, tellingly, output-only post-processing reduces the operating-point gap while leaving ABROCA untouched. All observed disparities were modest, below our pre-specified 0.08 minimum effect of interest, so we frame them as consistent, stage-localised patterns rather than confirmed effects. BPP reframes fairness auditing as an accounting problem across stages, showing practitioners where a fix will actually pay off.