Hiding a Gap Does Not Close It: The Price of Removing the Mean Group Contrast from Learned Models of Learning Dynamics

Sosthène Bidojesussi Hounzinhin, Ahmed Ratnani, Abdeladim El AkriarXiv 2026 · arXiv

TL;DR

This paper demonstrates that removing group information from student models in adaptive tutoring—a common "fairness" tactic—often incurs a high cost in predictive accuracy without actually helping students. The authors introduce a theoretical "price of erasure" and a three-stage audit tool to quantify this cost, arguing that fairness is better achieved through transparent policy decisions rather than by blinding the underlying models.

What Problem Does This Paper Address?

Adaptive tutoring systems use machine learning to predict a student's knowledge and decide which exercise or hint to provide next. However, the historical data used to train these models often reflects existing societal inequalities (e.g., gaps in school resources or socioeconomic status).

A popular remedy in the "fairness" literature is concept erasure: mathematically removing the "group signal" from the model so it cannot distinguish between, for example, students from wealthy vs. poor neighborhoods. The authors argue this is a contestable goal. If two groups do not perform equally in reality, a model forced to predict equal means is fundamentally wrong for at least one of them. Blinding the model doesn't fix the underlying inequality; it just makes the tutor less effective at identifying who actually needs help, potentially leading to "levelling down"—closing the gap by making everyone perform worse.

Methods

The paper combines theoretical derivations with empirical audits:

  1. The Price of Erasure: The authors derive a closed-form expression for the minimum extra error incurred when a model is forced to ignore group differences. For the best possible predictor, this cost is $p_0p_1\Delta^2/R^2$, where $p$ represents group shares, $\Delta$ is the predicted gap, and $R^2$ is how much the student's state explains their group membership.
  2. Three-Stage Audit: To make this practical, they propose a protocol:
  3. Fit: Train the predictor and a group model on training data.
  4. Announce: Estimate and "announce" the expected cost of erasure on validation data.
  5. Test: Verify if the gap reduction actually transfers to independent test units.
  6. Learning Dynamics: They use Koopman operators and Extended Dynamic Mode Decomposition (EDMD) to model how student knowledge evolves over time, rather than looking at static snapshots.
  7. Policy Design: They introduce a "floor-constrained" controller for tutoring. Instead of blinding the model, this controller uses a group-aware model but follows a policy that ensures no group’s knowledge falls below a certain "floor" while attempting to narrow the gap.

Datasets

The authors evaluated their approach on a mix of synthetic and large-scale real-world educational datasets:

  • ASSISTments 2012–2013: Data from an online math platform, using school-level performance as the group variable.
  • Eedi: Data from the NeurIPS 2020 Education Challenge, using "PremiumPupil" (eligibility for financial support) as a socioeconomic group indicator.
  • OULAD (Open University Learning Analytics Dataset): Distance learning data categorized by area deprivation, disability, and gender.
  • Tutoring Simulator: A custom-built simulator to test how different policies (like erasure vs. floors) affect long-term student outcomes.

Key Insights and Ideas

  • Alignment is the real killer: The "price" of erasing group information is highest when the group signal is aligned with the learning signal. In educational data, the features that tell you a student is "behind" are often the same features that predict their next answer. Erasing the group often erases the very information needed to teach effectively.
  • Erasure is a "Budget Effect": In simulations, much of the harm from erased models came from the fact that the blinded model underestimated the effect of help, leading the tutor to provide less assistance overall.
  • Blinding $\neq$ Fairness: On real data (Eedi), erasing the hidden state of a neural network still left 43% of the gap because the signal "leaked" through other inputs like the specific skill being tested.
  • Floors over Blinders: The authors show that you can narrow gaps without "levelling down" by using group-aware models paired with policies that explicitly protect the progress of the group that is behind.

Key Results

  • Accuracy Loss: On the OULAD dataset, the error of the "blinded" projected model was 7 to 11 times higher than the original model.
  • Audit Reliability: In the Eedi study, the three-stage audit successfully predicted the cost of erasure (an extra Brier score of ~0.021) and correctly identified when gap reductions would transfer to new classes.
  • Policy Failure: In the simulator, forcing the predicted gap to zero closed the gap in 23 of 24 settings but did so by lowering the knowledge of at least one group (levelling down).
  • The "Price" Formula: The theoretical formula for the price of erasure was found to be highly accurate on training data and a close approximation on held-out real-world students.

Limitations and Future Work

  • Binary Groups: The current mathematical framework focuses on binary group labels (e.g., Group 0 vs. Group 1). Extending this to continuous or multi-group attributes is a next step.
  • Observational Limits: The real-world studies are based on historical "replay" data. While they show how predictions fail, they cannot perfectly predict how a real student would react to a different tutoring policy in real-time.
  • Model-Based Control: The authors suggest that future work should focus on developing controllers that outperform simple "top-up" help strategies while maintaining the "floors" for group knowledge.

For more details, you can find the full paper under the title "Hiding a Gap Does Not Close It" on arXiv (2611.16100v1).

Summary updated

Abstract

Adaptive tutors use models learned from student data, and a common fairness remedy erases the group signal from these models. We turn the cost of this remedy into an audit tool. Building on mean-parity regression, we give the minimum extra error, relative to the best predictor, of removing the mean group contrast from a predictor that cannot use the group, $p_0p_1Δ^2/R^2$, and the exact cost of any correction built from an estimated group model, which, for the best predictor, is never below this minimum. A three-stage audit (fit on training units, announce the cost on validation units, test on independent units) estimates this cost before the correction is used. In a controlled study, the price can be large even when the state predicts the group poorly. On three real data sets, erasing the group from a learned model of learning dynamics predicts almost no gap where one is observed. In a selected knowledge-tracing network, erasing the hidden state leaves most of the gap; a correction of the prediction removes it at an extra Brier score of 0.021 on average. On Eedi, with the 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; on new ASSISTments schools the audit abstains. In a tutoring simulator, the objective and the help budget matter more than the erasure, and floors on each group's knowledge separate narrowing a gap from levelling down. Hiding a gap in a model does not close it.
BibTeX
@misc{hounzinhin2026hiding,
  author = {Sosthène Bidojesussi Hounzinhin and Ahmed Ratnani and Abdeladim El Akri},
  title = {{Hiding a Gap Does Not Close It: The Price of Removing the Mean Group Contrast from Learned Models of Learning Dynamics}},
  year = {2026},
  eprint = {2611.16100},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16100},
}
Figure