Pith. sign in

REVIEW 2 major objections 6 minor 35 references

When the Past Misleads: Rethinking Training Data Expansion Under Temporal Distribution Shifts

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding older data can hurt predictions; concept shift is the culprit

desk verdict The simulation is clean and useful; the empirical concept-shift metric needs validation before the case study can carry the load. read the letter →

arxiv 2509.01060 v2 pith:DRO63STM submitted 2025-09-01 cs.CY

Chengyuan Yao , Yunxuan Tang , Christopher Brooks , Rene F. Kizilcec , Renzhe Yu This is my paper
classification cs.CY
keywords temporaldistributionshiftconceptcovariateexpandingtrainingwindowalgorithmicfairnessintersectionalitystudentretentionpredictionAUCgap
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper challenges the default assumption that training predictive models on more historical data is always better. It argues that when the world is changing, expanding the training window backward can hurt performance, and the damage comes mostly from concept shift—changes in how features relate to outcomes—rather than covariate shift, changes in the features themselves. The argument is built from controlled simulations that decouple the two shifts and from a case study predicting first-year retention using 12 years of records from 23 community colleges. The paper also argues that fairness degrades when different demographic groups experience unequal concept shift, and that intersectional groups' experiences do not reduce to a simple sum of single-group effects. The practical stakes are that model builders should weigh temporal relevance, not just data volume, when choosing how much history to feed a model.

What carries the argument

The central object is the two-part decomposition of temporal distribution shift into covariate shift (changes in the marginal feature distribution P(X)) and concept shift (changes in the conditional outcome distribution P(Y|X)). The paper measures these with oracle Jensen–Shannon divergence on ground-truth coefficients in simulation, and in real data with a kNN-based estimate of P(Y|X) in PCA-reduced feature space, residualized against covariate shift using a random forest. Linear regressions (Equation 1 for performance, Equation 2 for fairness) attribute changes in AUC and in the AUC gap across groups to these shift metrics. The simulation framework is the controlled instrument that decoupl

What would settle it

Run a semi-synthetic experiment on the same college data: take a fixed test year, inject controlled concept shift by reweighting labels within feature neighborhoods while holding P(X) fixed, and check whether the residualized kNN concept-shift score tracks the injected amount. A second check injects only covariate shift and verifies that the residualized score stays flat; if either check fails, the central attribution of degradation to concept shift collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that under expanding historical training windows, concept shift—not covariate shift—is the primary driver of model performance degradation. In simulations, the AUC stays flat under covariate shift alone but declines as concept shift is introduced, and regression analyses show a consistently negative, statistically significant association between concept shift and performance. In the empirical student-retention setting, concept shift again predicts performance loss, especially in the largest training tertile (beta = -0.5412, p = .003), while covariate shift has small and inconsistent effects. For fairness, the gap in concept shift across demographic groups signifi

Load-bearing premise

The empirical concept-shift score is assumed to isolate true changes in P(Y|X) after regressing the kNN-based measure on covariate shift; if covariate shift distorts the kNN neighborhoods or the random-forest residualization absorbs genuine concept shift, the measured concept shift is confounded and the paper's causal interpretation weakens.

Editorial extensions

If this is right

  • Longer training windows are not universally beneficial; in settings with concept shift, they can reduce AUC rather than improve it.
  • Concept shift, not covariate shift, should be the primary monitored quantity when deciding whether to include older historical data.
  • Unequal concept shift across demographic groups predicts larger AUC gaps, so fairness auditing should track group-specific shift patterns over time.
  • Intersectional fairness cannot be extrapolated from single-group analyses; aligned versus opposing shift directions can reverse the direction of disparity.
  • Practitioners can treat training-window selection as a data-curation decision, potentially saving storage and compute while preserving or improving performance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension would test whether the residualized kNN concept-shift measure is a valid instrument: in semi-synthetic data, inject only concept shift while holding P(X) fixed and check whether the measure tracks the injection, and vice versa.
  • The paper's framing suggests that selectively dropping historical years with large concept shift could improve group fairness without adding fairness constraints—an empirical prediction the paper does not test.
  • The AUC-gap analysis could be generalized to other fairness metrics such as equalized odds or calibration gaps; if concept-shift gaps drive those too, the mechanism would be broader than ranking-based disparity.
  • The non-additivity result implies that auditing only single protected attributes may miss harms that emerge at intersections, so longitudinal subgroup monitoring may be necessary in practice.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This paper asks whether expanding historical training windows improves or harms predictive performance and fairness under temporal covariate and concept shift. The authors construct a 50-period simulation with logistic regression outcomes and four shift scenarios, plus two group-shift fairness scenarios, and regress AUC on covariate and concept shift (Eq. 1) and AUC Gap on inter-group shift gaps (Eq. 2). They find that concept shift has a significant negative association with AUC and that inter-group concept-shift gaps predict larger AUC gaps. An empirical case study on first-year retention across 23 minority-serving community colleges (2010-2021) uses a kNN-based, residualized concept-shift score. The full-sample regression on Eq. 1 is null; after a post-hoc tertile split by training-set size, concept shift is negatively significant only in the largest tertile (β=-0.5412, p=.003), and concept-shift gaps predict AUC gaps (β=0.5558, p=.002). The paper concludes that concept shift, not covariate shift, is the key driver of performance degradation and fairness loss when training windows expand.

Significance. If the central claim holds, the paper has clear practical value: it challenges the 'more data is better' assumption and offers a diagnostic lens for expanding-window training. Its strengths include a clean, oracle-based simulation that decouples covariate and concept shift, reproducible code, a large administrative dataset, and explicit attention to intersectional fairness. The simulation evidence is internally coherent. However, the empirical pillar is conditional on an unvalidated concept-shift estimator and on a post-hoc subgroup analysis, and the causal language outruns the observational regressions. The contribution is promising but needs validation before the broad conclusions can be accepted.

major comments (2)
  1. [Problem Setup and Methods, Measuring Distribution Shift (Concept Shift)] The empirical kNN+residualization concept-shift metric is never validated against ground truth. kNN neighborhoods are distorted by changes in P(X), and the random-forest residualization on the aggregate CovShift_unified cannot remove local interactions between covariate and concept shift. Since the simulation has an oracle concept-shift score, the same pipeline should be run on simulated data to show that the residualized score recovers the oracle. Without such validation, the significant coefficients in Eq. 1 (largest tertile) and Eq. 2 (β=0.5558, p=.002) could be artifacts of residual confounding. The paper's own limitations paragraph concedes sensitivity to feature scaling, high dimensionality, and sparsity.
  2. [Empirical Case Study, Results (Predictive Performance)] The full-sample regression on Eq. 1 shows no statistically significant main effects. The central empirical claim rests on a post-hoc tertile split by training-set size, in which only the largest tertile yields β=-0.5412 (p=.003). No multiple-testing correction or pre-specification is provided, and the phrase 'motivated by this consideration' indicates exploration. This does not, by itself, support the conclusion that concept shift is a key contributor to performance degradation. Report the full model, all tertile coefficients with confidence intervals, and a robustness check treating training size as a continuous moderator.
minor comments (6)
  1. [Eq. 1 and Eq. 2] Define the unit of analysis (school-window) and report sample sizes, standard errors, and fixed-effect estimates. Also clarify the notation I_emp·γ_j[i]; if the indicator is 1 only for empirical observations, make explicit that simulation units do not receive fixed effects.
  2. [Figures 4-6] The y-axis scales are not directly comparable across panels and metrics; state this in each caption and consider using standardized scales for at least one panel to ease visual comparison.
  3. [Discussion and Conclusion] The abstract and discussion use causal language ('drives', 'compromised') for empirical associations. Temper this language or state the identification assumptions needed for a causal interpretation, especially because the shift metrics are estimated from the same observational data used to evaluate model outcomes.
  4. [Simulation Study, Scenario (A)] The text says 'performance gains from longer training windows' in the no-shift scenario, but the reported result is 'largely flat with a small improvement at the beginning.' Align the wording with the figure.
  5. [General] Several references to 'Appendix' are not self-contained in the manuscript; ensure the appendix is stable and available with the submission, and include the full regression tables in the appendix.
  6. [Methods, Covariate Shift] The binary covariate-shift score averages absolute proportion differences and Cramér's V with equal weights; justify this aggregation choice or show robustness to alternative weights.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the simulation is a designed experiment with oracle concept-shift defined from data-generating coefficients, and the empirical concept-shift metric is estimated independently of the evaluated model.

full rationale

The paper's derivation chain is not circular. The simulation study generates data from explicit generative processes (logistic models with time-varying coefficients) and then evaluates a separately trained logistic model; the oracle concept-shift metric compares the true conditional label distributions under the training and test coefficients, while AUC is a consequence of the trained model's fit. These are distinct quantities, and the regression of AUC on the oracle metric is a designed experiment demonstrating a causal mechanism, not a tautology. In the empirical study, concept shift is estimated nonparametrically with kNN and then residualized on covariate shift using a random forest. Although this metric may be confounded by covariate shift (a limitation the paper acknowledges), it is not defined in terms of the model's own AUC or predictions; the paper explicitly avoids model-based drift detectors for this reason. The residualization is a standard two-step statistical control: the residual concept-shift score is orthogonal to covariate shift, and including it in the regression (Eq. 1) is equivalent to estimating the partial effect of raw concept shift controlling for covariate shift. No fitted parameter is later renamed as a prediction; the shift metrics are independent variables used to explain the outcome. The self-citations (e.g., Gardner et al. 2023 for the AUC Gap metric, and prior educational modeling work) provide metric choices and related work, but they are not load-bearing uniqueness theorems or ansatz smuggled in via citation; the metrics are defined explicitly in the paper. The absence of an external validation of the kNN+residualization pipeline is a validity limitation, not circularity, because the pipeline does not use the target outcome as an input. Overall, the central claims rest on a controlled simulation and an empirical regression, neither of which reduces by construction to its own inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the simulation scenario design (linear coefficient interpolation, drift rates) and on the validity of the kNN residualization as a measure of concept shift in the empirical data. No new entities are introduced.

free parameters (4)
  • k in kNN concept shift estimator = not stated
    The number of neighbors used to estimate conditional probabilities affects the concept shift metric; no sensitivity analysis is reported.
  • dPCA (number of PCA components retained) = not stated
    The dimensionality reduction for both covariate and concept shift metrics is chosen by hand; no criterion is given in the text.
  • Random forest hyperparameters for residualization = not stated
    The function f in ConceptShiftJS = f(CovShiftunified) + epsilon is a random forest; hyperparameters are not specified.
  • Simulation drift parameters (delta_j, rho_j, beta(0), beta(1), alpha_t) = in code/appendix
    These parameters define the shift scenarios; conclusions about relative impact of concept vs covariate shift may depend on their magnitudes.
assumptions (4)
  • domain assumption Logistic regression is the correct model class for the data-generating process and for retention prediction.
    The simulation generates data from a logistic model and trains logistic regression, so model-format mismatch is absent; in the empirical study, logistic regression is assumed to be a reasonable model for retention.
  • ad hoc to paper The kNN-based concept shift estimate after residualization reflects true changes in P(Y|X), not artifacts of covariate shift or feature scaling.
    Section 'Measuring Distribution Shift': the residualization uses a random forest fit on the same data, which is a modeling choice that must be valid for the empirical associations to be unbiased.
  • ad hoc to paper The simulation's linear interpolation beta(t) = (1 - alpha_t)beta(0) + alpha_t beta(1) captures realistic temporal concept shift.
    Used to generate scenarios; other shift dynamics could yield different performance-fairness relationships.
  • standard math Standard statistical machinery (KS test, Cramer's V, PCA, JS divergence, AUC, linear regression) is correctly applied.
    Used throughout the measurement and analysis pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When the Past Misleads: Rethinking Training Data Expansion Under Temporal Distribution Shifts." pith.science (2026). https://pith.science/paper/DRO63STM

@misc{pith2026250901060,
  author       = {Pith},
  title        = {Pith review of: When the Past Misleads: Rethinking Training Data Expansion Under Temporal Distribution Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DRO63STM}},
  note         = {Machine review of arXiv:2509.01060}
}
read the original abstract

Predictive models are typically trained on historical data to predict future outcomes. While it is commonly assumed that training on more historical data would improve model performance and robustness, data distribution shifts over time may undermine these benefits. This study examines how expanding historical data training windows under covariate shifts (changes in feature distributions) and concept shifts (changes in feature-outcome relationships) affects the performance and algorithmic fairness of predictive models. First, we perform a simulation study to explore scenarios with varying degrees of covariate and concept shifts in training data. Absent distribution shifts, we observe performance gains from longer training windows though they reach a plateau quickly; in the presence of concept shift, performance may actually decline. Covariate shifts alone do not significantly affect model performance, but may complicate the impact of concept shifts. In terms of fairness, models produce more biased predictions when the magnitude of concept shifts differs across sociodemographic groups; for intersectional groups, these effects are more complex and not simply additive. Second, we conduct an empirical case study of student retention prediction, a common machine learning application in education, using 12 years of student records from 23 minority-serving community colleges in the United States. We find concept shifts to be a key contributor to performance degradation when expanding the training window. Moreover, model fairness is compromised when marginalized populations have distinct data distribution shift patterns from their peers. Overall, our findings caution against conventional wisdom that "more data is better" and underscore the importance of using historical data judiciously, especially when it may be subject to data distribution shifts, to improve model performance and fairness.

Figures

Figures reproduced from arXiv: 2509.01060 by the authors.

Figure 1
Figure 1. Model performance and data distribution shift metrics across training window sizes. Each panel shows how AUC [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Subgroup model performance and shift metrics across training window sizes. Each panel presents AUC and shift [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Subgroup AUC and fairness gap under concept shift in same and opposite directions. Each panel shows subgroup [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: AUC trends by training start year for selected col [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Covariate shift and concept shift by training start year for selected colleges. Magnitudes are computed between each [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Subgroup AUC and concept shift by training start year for one college. The left panel shows AUC scores for four [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Azarkesht, M.; and Afsari, F. 2022. Instance reweighting and dynamic distribution alignment for domain adaptation. Journal of Ambient Intelligence and Humanized Computing, 13: 4967--4987

    work page 2022

  4. [4]

    Barocas, S.; Hardt, M.; and Narayanan, A. 2023. Fairness and Machine Learning: Limitations and Opportunities. The MIT Press. ISBN 9780262048613. Published December 19, 2023

    work page 2023

  5. [5]

    Beaulac, C.; and Rosenthal, J. S. 2019. Predicting University Students’ Academic Success and Major Using Random Forests. Research in Higher Education, 60(7): 1048--1064

    work page 2019

  6. [6]

    Berens, J.; Schneider, K.; Görtz, S.; Oster, S.; and Burghoff, J. 2019. Early Detection of Students at Risk: Predicting Student Dropouts Using Administrative Student Data from German Universities and Machine Learning Methods. Journal of Educational Data Mining, 11(3): 1--41

    work page 2019

  7. [7]

    Bifet, A.; Holmes, G.; and Pfahringer, B. 2010. Leveraging Bagging for Evolving Data Streams. In Balc \'a zar, J. L.; Bonchi, F.; Gionis, A.; and Sebag, M., eds., Machine Learning and Knowledge Discovery in Databases, 135--150. Berlin, Heidelberg: Springer Berlin Heidelberg. ISBN 978-3-642-15880-3

    work page 2010

  8. [8]

    A.; Castleman, B

    Bird, K. A.; Castleman, B. L.; and Song, Y. 2024. Are algorithms biased in education? Exploring racial bias in predicting community college student success. Journal of Policy Analysis and Management. First published: January 31, 2024

    work page 2024

Show all 35 references
  1. [9]

    Broby, D. 2022. The use of predictive analytics in finance. The Journal of Finance and Data Science, 8: 145--161

    2022

  2. [10]

    T.; Namkoong, H.; and Yadlowsky, S

    Cai, T. T.; Namkoong, H.; and Yadlowsky, S. 2023. Diagnosing Model Performance Under Distribution Shift. arXiv:2303.02011

    2023 arXiv

  3. [11]

    Chang, L.; and Krosnick, J. 2009. National Surveys Via RDD Telephone Interviewing Versus the Internet: Comparing Sample Representativeness and Response Quality. Public Opinion Quarterly, 73(4): 641--678

    2009

  4. [12]

    Csisz \'a r, I. 1975. I-Divergence Geometry of Probability Distributions and Minimization Problems. The Annals of Probability, 3(1): 146--158

    1975

  5. [13]

    R.; Ahsan, H.; Lakkimsetti, M.; Fatima, M.; Doshi, D.; Sadhu, K.; and Hassan, M

    Dixon, D.; Sattar, H.; Moros, N.; Kesireddy, S. R.; Ahsan, H.; Lakkimsetti, M.; Fatima, M.; Doshi, D.; Sadhu, K.; and Hassan, M. J. 2024. Unveiling the Influence of AI Predictive Analytics on Patient Outcomes: A Comprehensive Narrative Review. Cureus, 16(5): e59954

    2024

  6. [14]

    Gama, J.; Medas, P.; Castillo, G.; and Rodrigues, P. 2004. Learning with Drift Detection. In Bazzan, A. L.; and Labidi, S., eds., Advances in Artificial Intelligence--SBIA 2004, volume 3171 of Lecture Notes in Computer Science, 286--295. Springer, Berlin, Heidelberg

    2004

  7. [15]

    Gardner, A. 2022. Persistence and Retention: Fall 2020 Beginning Postsecondary Student Cohort. Technical report, National Student Clearinghouse Research Center, Herndon, VA

    2022

  8. [16]

    Gardner, J.; Yu, R.; Nguyen, Q.; Brooks, C.; and Kizilcec, R. 2023. Cross-Institutional Transfer Learning for Educational Models: Implications for Model Performance, Fairness, and Equity. In 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’23. ACM

    2023

  9. [17]

    R.; Orona, G

    Glandorf, D.; Lee, H. R.; Orona, G. A.; Pumptow, M.; Yu, R.; and Fischer, C. 2024. Temporal and Between-Group Variability in College Dropout Prediction. In Proceedings of the 14th Learning Analytics and Knowledge Conference, 486--497

    2024

  10. [18]

    M.; Rasch, M

    Gretton, A.; Borgwardt, K. M.; Rasch, M. J.; Sch \"o lkopf, B.; and Smola, A. 2012. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13(25): 723--773

    2012

  11. [19]

    Hu, Q.; and Rangwala, H. 2020. Towards Fair Educational Data Mining: A Case Study on Detecting At-risk Students. In Proceedings of the 13th International Conference on Educational Data Mining (EDM), 431--437

    2020

  12. [20]

    F.; and Lee, H

    Kizilcec, R. F.; and Lee, H. 2022. Algorithmic Fairness in Education. In Ethics in Artificial Intelligence in Education. Routledge

    2022

  13. [21]

    Klinkenberg, R.; and Joachims, T. 2000. Detecting Concept Drift with Support Vector Machines. In Proceedings of the Seventeenth International Conference on Machine Learning, ICML '00, 487–494. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc. ISBN 1558607072

    2000

  14. [22]

    intersectionally fair

    Kong, Y. 2022. Are “intersectionally fair” ai algorithms really fair to women of color? A philosophical analysis. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 485--494. Association for Computing Machinery

    2022

  15. [23]

    M.; and Loog, M

    Kouw, W. M.; and Loog, M. 2019. An introduction to domain adaptation and transfer learning. arXiv:1812.11806

    2019 arXiv

  16. [24]

    L.; Khoshgoftaar, T

    Leevy, J. L.; Khoshgoftaar, T. M.; Bauder, R. A.; and Seliya, N. 2019. The Effect of Time on the Maintenance of a Predictive Model. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), 1891--1896

    2019

  17. [25]

    Lu, W.; Wang, J.; Li, H.; Chen, Y.; and Xie, X. 2022. Domain-invariant Feature Exploration for Domain Generalization. arXiv:2207.12020

    2022 arXiv

  18. [26]

    W.; and Wang, J

    Marsaglia, G.; Tsang, W. W.; and Wang, J. 2003. Evaluating Kolmogorov's Distribution. Journal of Statistical Software, 8(18): 1--4

    2003

  19. [27]

    Mehrabi, N.; Morstatter, F.; Saxena, N.; Lerman, K.; and Galstyan, A. 2021. A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys (CSUR), 54(6): 1--35

    2021

  20. [28]

    Quiñonero-Candela, J.; Sugiyama, M.; Schwaighofer, A.; and Lawrence, N. D., eds. 2009. Dataset Shift in Machine Learning. Neural Information Processing. The MIT Press. ISBN 9780262545877. Published June 7, 2022

    2009

  21. [29]

    Sghir, N.; Adadi, A.; and Lahmer, M. 2023. Recent advances in Predictive Learning Analytics: A decade systematic review (2012--2022). Education and Information Technologies, 28: 8299--8333

    2023

  22. [30]

    Tang, Y.; Harvey, E.; Yao, C.; Yu, R.; Kizilcec, R.; and Brooks, C. 2025. Understanding Predictive Models of Student Success with a Multiverse Analysis. In Mills, C.; Alexandron, G.; Taibi, D.; Bosco, G. L.; and Paquette, L., eds., Proceedings of the 18th International Confere...

    2025

  23. [31]

    Department of Education

    U.S. Department of Education . 2025. Community College Facts at a Glance

    2025

  24. [32]

    V.; and Russakovsky, O

    Wang, A.; Ramaswamy, V. V.; and Russakovsky, O. 2022. Towards Intersectionality in Machine Learning: Including More Identities, Handling Underrepresentation, and Performing Evaluation. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAc...

    2022

  25. [33]

    Xu, Y.; and Wilson, K. 2021. Early Alert Systems During a Pandemic: A Simulation Study on the Impact of Concept Drift. In LAK21: 11th International Learning Analytics and Knowledge Conference, LAK21, 504–510. New York, NY, USA: Association for Computing Machinery. ISBN 9781450389358

    2021

  26. [34]

    Yao, C.; Cortez, C.; and Yu, R. 2025. Towards Fair and Privacy-Aware Transfer Learning for Educational Predictive Modeling: A Case Study on Retention Prediction in Community Colleges. In Proceedings of the 15th International Learning Analytics and Knowledge Conference, LAK '25...

    2025

  27. [35]

    Yu, R.; Lee, H.; and Kizilcec, R. F. 2021. Should College Dropout Prediction Models Include Protected Attributes? In Proceedings of the Eighth ACM Conference on Learning @ Scale, L@S '21, 91–100. New York, NY, USA: Association for Computing Machinery. ISBN 9781450382151

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.