Accuracy Is Not Enough: A Cross-Architecture Audit of Demographic Bias in Deep Knowledge Tracing
Dang Quang Minh, Nguyen Dung Son, Nguyen Huu Loi, Truong Viet Vu, Nguyen Thai Anh
cs.LG
Quick verdict
First cross-architecture audit of deep-KT demographic bias; statistically careful and honest, but headline p-values sit at the permutation floor.
Summary
The paper presents what the authors describe as the first cross-architecture demographic fairness audit of deep knowledge tracing, covering four models (DKT, DKVMN, SAKT, AKT) trained under three regimes (standard, reweighting, adversarial debiasing) on two public datasets with demographic metadata: Eedi (15.9M interactions, 118,971 students) and OULAD (166,875 submissions, 19,822 students). Audits use ABROCA with student-level bootstrap confidence intervals, permutation tests, and Benjamini-Hochberg correction. Socioeconomic (free-school-meal) bias is significant for all four architectures on Eedi (ABROCA 0.018-0.023, p<0.005), while gender bias is significant on OULAD for three of four architectures after multiplicity correction but null on Eedi. AKT, the most accurate model (AUC 0.807 vs 0.757-0.765), shows the largest socioeconomic bias; ablating its Rasch embeddings removes both the accuracy gain and the excess bias, and a paired bootstrap supports the difference (p<=0.002). Reweighting and adversarial debiasing leave ABROCA essentially unchanged in every accuracy-preserving configuration, with mechanism checks verifying the adversary works. Limitations: headline p-values sit at the 200-permutation resolution floor, the OULAD gender effect flips across correctness-threshold definitions (40/60/70), 71% of Eedi students lack SES labels, and no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Every deep KT architecture shows significant socioeconomic (FSM) bias on Eedi (ABROCA 0.018-0.023, p<0.005), while gender bias is significant on OULAD for three of four architectures after BH correction and null on Eedi.
- AKT's item-level Rasch embeddings simultaneously buy about 4 AUC points of accuracy and the largest socioeconomic bias; ablating them removes both, verified by paired bootstrap (p<=0.002) and three-seed stability.
- Reweighting and adversarial debiasing leave ABROCA essentially unchanged (deltas <=0.0015, p>=0.27) in every accuracy-preserving configuration, with the adversarial mechanism verified in both directions.
- Fairness conclusions do not transfer across datasets, attributes, or correctness-threshold operationalizations, so KT audits must be run per dataset, per attribute, and per skill.
Strengths
Genuine gap, comprehensive grid
Prior KT fairness work is BKT-centric (Zambrano 2024, Stinar 2025, Barrett 2024); this is the first 4-architecture x 3-regime x 2-dataset audit of deep KT (56 main runs, 72 total), per Section 1 and Fig. 1.
Statistical machinery answering a known critique
Section 3.3 operationalizes Borchers-Baker (2025): student-level bootstrap CIs (200 resamples), permutation tests, BH correction across the 44-test grid, and paired student-level bootstraps (1,000 replicates) on identical resamples; explicitly avoids inferring significance from bootstrap intervals on a non-negative statistic.
Controlled ablation attributing accuracy and bias to one component
Mitigation null verified, not assumed
Honest self-auditing of design choices
Weaknesses
Headline evidence capped at the permutation resolution floor
OULAD gender headline is threshold-fragile
No code or artifact availability stated
Coverage and external-validity gaps
Fairness-metric scope leaves decision-relevant gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper claims the first cross-architecture demographic fairness audit of deep knowledge tracing: four architectures (DKT, DKVMN, SAKT, AKT) x three training regimes (standard, reweighting, adversarial) x two public datasets (Eedi, 15.9M interactions / 118,971 students; OULAD, 166,875 submissions / 19,822 students), audited with ABROCA, student-level bootstrap CIs, permutation tests, and Benjamini-Hochberg (BH) correction. Three headline claims: (i) socioeconomic (FSM) bias is significant for every architecture on Eedi (ABROCA 0.018-0.023, p<0.005) while gender bias appears on OULAD (3/4 architectures after BH) but not Eedi; (ii) the most accurate model, AKT (AUC 0.807 vs 0.757-0.765), is the most biased (SES ABROCA 0.023 vs 0.018; paired delta +0.005, p<=0.002), with a Rasch-embedding ablation removing both the accuracy gain (to 0.766) and the excess bias (to 0.018); (iii) reweighting and adversarial debiasing leave ABROCA essentially unchanged in every accuracy-preserving configuration. Claims are carefully scoped: no downstream-harm assertion, no post-processing claims.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ABROCA in knowledge tracing fairness audits?
ABROCA (absolute between-ROC area) measures fairness by integrating the absolute gap between the ROC curves of two demographic groups across all classification thresholds. Unlike comparing accuracy or AUC at a single point, it captures systematic performance differences over the full score range. This paper reports ABROCA with student-level bootstrap confidence intervals and permutation p-values, because small samples and class imbalance are known to inflate the metric by chance.
Do deep knowledge tracing models discriminate against certain student groups?
This audit of DKT, DKVMN, SAKT, and AKT found small but statistically significant socioeconomic bias on Eedi (models predict free-school-meal students slightly worse, ABROCA about 0.02) and gender bias on OULAD for three of four models; direction and size depend heavily on dataset and attribute. The authors argue these are systematic, mitigation-resistant signals rather than demonstrated harms, and they do not claim downstream impact without a decision-layer study.
Can reweighting or adversarial debiasing fix fairness problems in knowledge tracing models?
In this study, no. Both in-training techniques left the ABROCA bias metric essentially unchanged (deltas at most 0.0015) in every configuration that preserved accuracy. The only apparent improvement came from an adversarial setting so strong it collapsed accuracy by 13 AUC points, and even that bias reduction was statistically indistinguishable from zero.
Which knowledge tracing model is most accurate, and is it the fairest?
AKT was the most accurate on Eedi (AUC 0.807 versus 0.757-0.765 for DKT, DKVMN, and SAKT), thanks to its Rasch item-difficulty embeddings, but it also showed the largest socioeconomic bias. Removing only the Rasch embeddings returned both accuracy and bias to the simpler models' levels, linking the accuracy gain and the fairness cost to the same component.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are available only as captions; the bootstrap distribution and per-skill density plots could not be visually inspected.
- Extraction may omit an appendix or supplement; a code-availability statement elsewhere in the source cannot be fully ruled out.
- The metadata abstract is truncated mid-sentence, but the full body text compensates completely.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.