Skip to main content
Back to timeline
arXivSource publication:

Erasing a group gap from a learning model does not close it: an audit prices the removal at about 0.021 Brier score on new Eedi classes

Synopsis

Adaptive tutors often pursue fairness by erasing the group signal from a learned model, but what is erased is the mean group contrast in predictions, not the real gap. The work gives a closed-form minimum extra error for this removal and a three-stage audit that fits on training units, announces the cost on validation units and tests on independent units. On three real data sets the erased model predicts almost none of an observed gap; on Eedi, with a knowledge-tracing network selected on separate units, the audit accepted the announced cost in all five splits, the measured cost fell inside the announced interval in all five, and the gap reduction transferred to new classes in four.

Source-provided article image: Hiding a Gap Does Not Close It: The Price of Removing the Mean Group Contrast from Learned Models of Learning Dynamics
arXiv

Interpretation

Removing the mean group contrast from a predictor that uses only the student state carries an extra error that can be computed in advance, and on real learning data that error is not small. Earlier practice projected the group direction out of a model representation or its dynamics so that the predicted gap fell near zero, without stating how much predictive accuracy the operation costs. On ASSISTments, Eedi and OULAD the projected model predicts a gap near zero while the observed gap on held-out units is +0.057, −0.085 and −0.048; rank-matched random directions and lowest-variance PCA cost far less than the group direction.

The price has a closed form: the variance of the best prediction times its squared correlation with the group signal, equivalently the product of group shares times the squared predicted gap divided by how well the state explains the group. The structure matches known mean-parity regression results; here it is applied to a predictor of the state alone, with two estimable equivalent forms and a bound. Across five real settings the formula matches the measured extra error on training data and stays within 0–13% of it on held-out students; gender is weakly aligned with the group signal (|ρ|=0.27) and cheap, while disability is strongly aligned (|ρ|=0.77) with a state that barely explains the group (R²=0.004) and costs about eight times more.

The audit announces the cost of the correction that can actually be built and abstains when the data do not support an estimate. Measuring the error on the units that fit the models is biased; an earlier version announced 0.79 for a population cost of 1.01. On Eedi, with the network selected on separate units, all five splits accepted the announced cost, the measured cost of 0.017–0.027 (mean 0.021) lay inside the announced interval in all five, and the gap reduction transferred to new classes in four; on new ASSISTments schools the audit abstained in all five splits.

Perspective

The conclusions rest on prediction comparisons on three observational data sets and on a purpose-built tutoring simulator, and do not evaluate a tutor used by real students; ASSISTments groups are retrospective school-performance groups, Eedi and OULAD use attributes recorded outside the data, and Eedi covers only the 29% of pupils whose support status is recorded.

The audit intervals are approximate: in the controlled study the nominal 95% interval contained the realized cost in 90–100% of accepted replicates, on Eedi the intervals were 1.0–2.0 times as wide as the cost, and the five splits share units rather than being five independent validations; when the group model misses part of the group signal, the announced cost is conservative as an estimate of the minimum, and the diagnostic cannot detect this.