acceptodds
Under review as a conference paper at ICLR 2027

Fairness by Design: Machine Learning and Interpretable Mortgage Lending

Abstract

Regulated lenders must explain denials and avoid disparate treatment in U.S. mortgages, yet flexible scoring can be opaque. We develop and study a less discriminatory and transparent machine-learning approach that embeds equalized-odds fairness in estimation and decomposes each decision into feature contributions suitable for adverse-action notices. Using HMDA applications, the model preserves performance while shrinking minority–nonminority error gaps and reweighting away from geographic proxies toward core underwriting signals (debt-to-income, loan-to-value). Supplementary analyses around underwriting cutoffs and lender-level tests indicate that the model lowers minority shortfalls at the margin and reduces within-lender disparities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Credit, Mortgages, Machine learning, Fairness, Interpretability, Regulations

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QuhrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a Fairness-Aware Additive Network (FAN) for mortgage lending. The aim is to reduce group disparities while still maintaining predictive performance and interpretability. It's an additive neural network that takes univariate features and sparse pairwise interactions. The training loss is a weighted sum of binary cross-entropy, a soft equalized odds (EO) penalty, and a 1-norm sparsity penalty. The authors also provide a theoretical analysis of the relationship between the soft EO penalty and the threshold-averaged EO penalty and derive a uniform convergence bound for the empirical EO penalty. The method is evaluated on the HMDA dataset.

Critical strengths

The paper is a fairly comprehensive empirical analysis of fairness-aware mortgage lending using large-scale HMDA data. In addition to standard predictive and fairness metrics, they examine within-lender variation, institutional underwriting thresholds, and robustness over time. Together, these findings provide useful evidence of the practical relevance of fairness regulation.

Critical weaknesses
  1. The methodological and theoretical novelty is very limited.

The proposed FAN architecture is essentially a re-implementation of GAMI-Net (Yang et al., 2021) with the same additive neural architecture, the same sparse pairwise interactions and the same three stage training process. The main methodological difference is the inclusion of a differentiable equalized odds (EO) penalty which has been studied before in the literature (e.g. Pfohl et al., 2021). The theoretical contributions are also very limited. Prop 1 is a standard integral identity. Prop 2 does not actually demonstrate that minimizing the EO penalty will control the threshold based fairness gap. Prop 3 is a standard uniform convergence bound for the soft EO penalty. There's essentially no new methodological or theoretical contribution here.

  1. The baselines in the comparisons are weak.

The evaluation is limited to logistic regression and unconstrained FAN with no comparison made to GAMI-Net or any of the other well known fairness aware learning methods. The experiments do show that adding the EO regularization to FAN does result in better fairness metrics, but they don't actually show that it is better than any of the other methods. That's a significant methodological gap.

Reference

  • Yang, Z., Zhang, A., & Sudjianto, A. (2021). GAMI-Net: An explainable neural network based on generalized additive models with structured interactions. Pattern Recognition, 120, 108192.

  • Pfohl, S. R., Foryciarz, A., & Shah, N. H. (2021). An empirical characterization of fair machine learning for clinical risk prediction. Journal of Biomedical Informatics, 113, 103621.

Questions for authors
  1. Methodological novelty.

The FAN architecture is very similar to GAMI-Net (Yang et al. 2021) in that it has an additive neural structure, sparse pairwise interactions, and a 3 stage training process. Can the authors be more specific about the architectural or algorithmic differences between FAN and GAMI-Net? Pfohl et al. (2021, eq 9-12) already introduce a differentiable EO regularizer based on the squared differences in group-conditional mean. For two protected groups, their formulation reduces to a weighted version of the penalty in FAN. Could the author clarify the novelty of their EO regularizer beyond the normalization and the use of predicted probabilities instead of log-probabilities?

  1. Theoretical contributions.

Prop 2 establishes Pmean≤D1+D0P_\mathrm{mean} \le D_1 + D_0, which is simply a consequence of the non-negativity of variance. For two protected groups we have Δy(τ)=R1,y(τ)−R0,y(τ)\Delta_y(\tau) = R_{1,y}(\tau) - R_{0,y}(\tau). Based on the uniform variance across groups assumption (line 1422-1426), it follows that Pmean=14∑y=01(Ew[Δy])2,P_{\mathrm{mean}}=\frac14\sum_{y=0}^{1} \left(\mathbb E_w[\Delta_y]\right)^2, and D1+D0=Pmean+14∑y=01Var⁡w(Δy).D_1+D_0=P_{\mathrm{mean}}+ \frac14\sum_{y=0}^{1}\operatorname{Var}_w(\Delta_y). The 14\frac{1}{4} factor changes if the group weights are unequal but the variance decomposition still hold.

Since the training objective is to minimize PmeanP_\mathrm{mean} and the variance term isn't directly under control then Prop 2 doesn't establish that minimizing the EO penalty will lead to reduced threshold averaged EO disparities. Can the authors clarify how this result supports their stated fairness guarantee? Are further assumptions required to control the variance term? More generally what are the non-trivial theoretical contributions of Props 1-3 beyond the standard integral identities, variance inequalities and uniform convergence arguments?

  1. Experimental comparisons.

The evaluation looks at FAN against logistic regression and unconstrained FAN. Are there any comparisons against GAMI-Net and other established fairness-aware learning methods? Given the substantial architectural similarities between FAN and GAMI-Net, FAN with λEO=0\lambda_\mathrm{EO} = 0 appears to approximate an unconstrained GAMI-Net. Could the authors clarify which differences between these models affect the performance? In particular, what evidence shows that FAN with EO regularization offers advantages over GAMI-Net with the EO regularization? Such comparisons would help establish the incremental value of the proposed method.

1 comment
  1. Contribution: FAN builds on additive neural modeling, as described in Section 3.1, p. 3, and Appendix D, pp. 18–22, but builds a custom framework for the empirical question of interest. Its principal modification is fairness-aware estimation while preserving exact additive score decompositions. Pfohl et al. provide a relevant antecedent for outcome-conditional mean-matching penalties. However, normalization and the choice of probabilities versus log-probabilities are not the basis of our contribution. The contribution is therefore the integrated framework and the economic evidence it enables. Sections 6–7, pp. 5–9, show how fairness adjustments reduce reliance on geographic proxies while preserving core underwriting signals, and how disparities decline within lenders even when approval capacity is fixed. The existence of the individual methodological ingredients does not establish these findings or explain how fairness interventions change model-implied mortgage allocation. Finally, the paper’s contribution should also be assessed in the financial-economic and regulatory context stated in Section 1, pp. 1–2: can mortgage decision rules reduce group disparities while retaining nonlinear predictive performance and providing transparent explanations of individual decisions?

  2. Theoretical contributions: Proposition 2 establishes a lower bound, not a guarantee that minimizing the soft penalty reduces threshold-averaged disparity. Appendix F, p. 26 already excludes a pointwise fairness guarantee. The role of Proposition 1 is to give the conditional mean probabilities a threshold-average interpretation and Proposition 3 concerns generalization of the soft penalty under the stated assumptions. A stronger guarantee for averaged disparity would require additional control of the variance remainder or direct regularization of the averaged disparity. Our empirical conclusions are supported by direct evaluation: Appendix H, Table 8, p. 35, reports that test-set threshold-weighted disparity falls from 0.00114 to 0.00046. Appendix I, Table 9, p. 37, reports smaller hard-decision error-rate gaps at matched approval rates, with additional threshold sweeps in Tables 10–11, pp. 38–39. These are empirical improvements in this application, rather than consequences of reversing Proposition 2.

  3. Experimental comparisons: The existing comparisons already address whether incorporating fairness into an interpretable mortgage model produces less disparate decisions while preserving predictive fit. Section 5, Table 1, p. 4, evaluates this within the FAN model family. Appendix I, Table 9, p. 37, matches overall approval rates and Section 7, Table 3, p. 9, fixes approval capacity within lenders. These results support the paper’s economic finding that candidate less discriminatory decision rules exist within a common public-information set. The manuscript does not report a separate comparison against other models trained with the identical EO penalty, and therefore does not intend to establish an architectural advantage over such models. Such comparisons would be required to claim superiority over alternative fairness algorithms, but they do not negate the documented fairness–fit trade-off or its economic interpretation.

Reviewer_c8MLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FAN, a neural additive model with sparse pairwise interactions, together with a smooth equalized-odds penalty. The method is evaluated on HMDA mortgage approval data using a binary minority group indicator. Compared with logistic regression, FAN achieves better AUC, while increasing the fairness penalty reduces the gaps in TPR and FPR with only a small loss in predictive performance. Its additive structure also makes it possible to provide feature level explanations.

Critical strengths
  1. The paper addresses an important and well-motivated problem at the intersection of fairness and interpretability in regulated lending.

  2. The model is interpretable by design, and the comparison of learned shape functions across different fairness penalty strengths provides useful insight into how the model changes as the fairness constraint becomes stronger.

  3. The empirical study is based on a large real-world lending dataset, and the paper also includes several robustness checks that strengthen the evaluation.

Critical weaknesses
  1. Limited novelty FAN appears to combine an existing interpretable additive architecture with an existing fairness regularization idea, but the paper does not sufficiently position itself relative to closely related work. The theoretical results also seem fairly standard: Propositions 1–3 are essentially applications of the layer-cake identity, Jensen’s inequality, and a Rademacher complexity bound.

    For interpretable additive models with interactions, several closely related papers should be discussed, including:

    • Agarwal et al., “Neural Additive Models: Interpretable Machine Learning with Neural Nets” (NeurIPS 2021).
    • Yang et al., “GAMI-Net: An Explainable Neural Network Based on Generalized Additive Models with Structured Interactions” (Pattern Recognition 2021). This paper already uses main-effect and interaction subnetworks, FAST-based interaction screening, heredity constraints, and a three-stage training procedure that are very similar to the components used in FAN, yet it is not cited.
    • Chen & Ye, “Monotonic Neural Additive Models: Pursuing Regulated Machine Learning Models for Credit Scoring” (arXiv 2022).

    Likewise, equalized-odds-style penalties and fairness constraints during training have been studied extensively, for example in:

    • Zafar et al., “Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatment” (WWW 2017).
    • Bechavod & Ligett, “Penalizing Unfairness in Binary Classification” (arXiv 2017).
    • Zhang et al., “Mitigating Unwanted Biases with Adversarial Learning” (AIES 2018).
    • Agarwal et al., “A Reductions Approach to Fair Classification” (ICML 2018).
    • Cruz et al., “FairGBM: Gradient Boosting with Fairness Constraints” (ICLR 2023).

    The paper therefore needs to explain much more clearly what is genuinely new beyond combining these existing ingredients.

  2. Narrow scope The current formulation only considers a binary sensitive attribute and a binary outcome. Equation (2) is explicitly defined for two groups, and all experiments are based on a single dataset. This is a fairly restrictive setting given that existing fairness methods already address multiple sensitive groups, continuous protected attributes, and non-binary outcomes. Relevant examples include:

    • Kearns et al., “Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness” (ICML 2018).
    • Mary et al., “Fairness-Aware Learning for Continuous Attributes and Treatments” (ICML 2019).
    • Agarwal et al., “Fair Regression: Quantitative Definitions and Reduction-Based Algorithms” (ICML 2019).
  3. Weak baselines and an unfair comparison The only baseline is an unregularized logistic regression, which is not sufficient to establish the benefit of the proposed method. The paper does not compare against established fairness methods such as Hardt et al., “Equality of Opportunity in Supervised Learning” (NeurIPS 2016), reduction-based approaches, adversarial debiasing, or Cruz et al., “FairGBM: Gradient Boosting with Fairness Constraints” (ICLR 2023). It also does not compare against interpretable models such as Agarwal et al., “Neural Additive Models: Interpretable Machine Learning with Neural Nets”, or Yang et al., “GAMI-Net: An Explainable Neural Network Based on Generalized Additive Models with Structured Interactions”, trained with the same fairness penalty.

    More importantly, comparing a nonlinear neural model with fairness regularization against a linear logistic regression without any fairness constraint mixes together the effects of model capacity and fairness regularization. A stronger evaluation would hold the model class fixed where possible, or compare methods at matched predictive performance or matched approval rates.

  4. Insufficient ablation and uncertainty analysis The paper does not provide ablations for several components of the method, including the interaction terms, gating mechanism, warm starts, clarity penalty, or staged training procedure. It is therefore difficult to tell which components are actually responsible for the observed behavior. In addition, the fairness penalty is only evaluated at λ∈0,1,2,3\lambda \in \\{0,1,2,3\\}, which gives a rather coarse view of the fairness–accuracy trade-off. The experiments also do not report repeated runs, random-seed variability, or confidence intervals.

Questions for authors
  1. Beyond incorporating an equalized-odds penalty, what are the main methodological contributions of FAN relative to GAMI-Net?

  2. How does FAN compare with existing fairness methods and interpretable baselines when predictive performance or approval rates are matched?

  3. Can the proposed framework be extended to multiple or intersectional sensitive groups, as well as to non-binary outcomes?

2 comments
  1. Comment 1: The novelty lies in the custom integrated framework and the economic evidence it enables by combining the additive structure with fairness regularization. This contribution addresses a substantive question in financial economics and lending regulation: can mortgage decision rules reduce group disparities while retaining nonlinear predictive performance and providing transparent explanations of individual decisions? As explained in Section 1, pp. 1–2, the search for less discriminatory alternatives must account jointly for predictive performance, fairness, and the ability to explain adverse decisions. To the best of our knowledge, this is the first mortgage-lending study to integrate equalized-odds regularization directly into an inherently interpretable nonlinear additive neural model and examine how this integration changes underwriting rules and within-lender disparities. Interpretability is central to the economic analysis: the exact decomposition of denial scores reveals how fairness adjustments change reliance on geographic proxies, repayment-capacity measures, and collateral characteristics. Sections 6–7, pp. 5–9, show that disparities can decline while preserving predictive fit to observed underwriting decisions, including when approval capacity is held fixed within lenders. The contribution therefore includes both the integrated framework and the economic evidence it enables. Propositions 1–3, Appendices F–G, pp. 26–34, characterize the chosen penalty using established mathematical tools; novel proof techniques are not the basis of the contribution. The manuscript already presents FAN within the additive-model framework (Section 3.1, p. 3).

  2. Comment 2: The empirical scope follows the economic question: mortgage underwriting involves a denial-versus-origination decision, as defined in Section 4, p. 4. Continuous-outcome methods therefore address a different prediction task. Moreover, the theory already accommodates multiple protected groups: Appendix G, p. 27, defines A∈GA\in\mathcal G and measures disparities across a finite group set. Appendix L, pp. 50–53, replicates the analysis using 2019 HMDA; Appendix M, p. 54, examines bank-originated jumbo applications; and Appendix N, pp. 55–56, evaluates subgroup disparities. These exercises assess robustness within the mortgage setting that motivates the paper.

  3. Comment 3: The recommendation to hold the model family fixed describes the central comparison already reported in the manuscript. Section 5, Table 1, p. 4, compares unconstrained FAN with fairness-regularized FAN, showing smaller disparities with essentially unchanged predictive performance. Appendix I, Table 9, p. 37, compares models at matched approval rates, while Section 7, Table 3, p. 9, fixes approval capacity within each lender. These comparisons directly address the economic question of whether lenders can reduce disparities by changing their scoring rules without expanding approval capacity. The finding does not depend on the logistic-versus-neural comparison. Broader benchmarking would address comparative superiority over other fairness algorithms, which the current results do not intend to establish.

  4. Comment 4: Table 1, p. 4, already examines the central ablation by removing the fairness penalty at λEO=0\lambda_{\mathrm{EO}}=0 and increasing its weight to 1,2,3.Appendix H, Table 8 and Figure 8, pp. 35–36, reports fairness diagnostics across data splits and variability from repeated evaluation subsamples. Appendix I, pp. 37–40, additionally examines decision-threshold sensitivity and matched approval rates. These analyses support the reported fairness improvements over the tested settings. Comprehensive component ablations, a denser regularization grid, and repeated training-seed experiments are additional analyses that the current manuscript does not focus on.

Thank you for the detailed response. Some points are clarified, but my main concerns remain.

  1. I agree that Table 1 isolates the effect of the penalty within FAN, and that Table 9 compares models at matched approval rates. However, there is still no comparison with established fairness methods, such as reductions, adversarial debiasing, or post-processing as in Hardt et al., "Equality of Opportunity in Supervised Learning" (NeurIPS 2016). There is also no comparison with interpretable models such as NAM or GAMI-Net trained with the same penalty. Showing that less discriminatory alternatives exist does not require FAN specifically, so it remains unclear what FAN adds over existing tools.

  2. Component ablations (interaction terms, gating, staged training), a finer grid of λEO\lambda_{\text{EO}}, and variability across training seeds are still missing. Variability across evaluation subsamples does not capture the randomness of neural-network training.

  3. I accept that the binary outcome follows from the mortgage application. However, the method and experiments remain restricted to a binary sensitive attribute and to HMDA data.

Finally, Proposition 2 provides only a lower bound and does not guarantee control of the threshold-averaged EO disparity. Second, the O(L/n)O(L/\sqrt{n}) rate in Proposition 3 does not follow from the current proof for learned subnetworks. As a result, the theoretical contribution is weaker than presented, and the contribution rests mainly on the empirical study, which is also limited.

Since no revised manuscript has been provided, I will keep my original score.

Reviewer_RqZWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces an in-processing fairness algorithm for EO called FAN - in particular, it is a generalized additive model with neural subnetworks and an EO fairness penalty with sparsity regularization. They evaluate their model on a large mortgage loaning dataset and interpret various results of their learned model, e.g., which predictors remain relevant or how the shape functions vary for the fair model. Their empirical results are supported with some theoretical results including out of sample generalization error and threshold based EO disparities.

Critical strengths

Their empirical analysis is thorough - for their chosen application, they investigate a number of results, like learned shape functions and alternative thresholds, and then translate this into economic terms. I also appreciate the effort to provide a theoretical justification for the fairness penalty, although I have concerns about the interpretation and completeness of these results.

Critical weaknesses

It appears that FAN combines methodological elements that have already been introduced in GAMI-Net (Yang et al., 2021) and Neural Additive Models with EO-regularized training (e.g., FNNC, Padala et al. 2020). The stated theoretical results also seem to have some flaws (see questions below), so, overall, the technical novelty appears limited.

On the empirical side, the implications of their analyses seems overstated. Since the labels represent historical lending decisions, EO measures disagreement with lenders, rather than fairness relative to creditworthiness (i.e., mismatching interpretations of the response Y); this also limits the justification in Appendix C. Consequently, downstream claims about less discriminatory lending and preserved underwriting quality are tenuous.

Questions for authors

Could the authors compare FAN with existing interpretable models trained using the same fairness penalty, such as GAMI-Net or FNCC?

For proposition 2, the main conclusion is P_{mean} \leq D_1+D_0. Control of the lower bound does not generally control the quantity above it, so how does this result justify the claim that reducing the penalty controls threshold-averaged EO disparity?

Lemma 3 in Appendix G.3 treats the component functions as a fixed finite dictionary, whereas FAN learns these functions and selects components from data (and is thus adaptive in nature). Could the authors provide a complete justification for the extension to learned subnetworks- in particular, the asserted O(L / \sqrt n) complexity rate?

1 comment
  1. Comparisons with other models: Our contribution concerns the integration of fairness and interpretability in mortgage lending, together with evidence on how fairness constraints change geographic-proxy use and within-lender disparities (Sections 6–7, pp. 5–9). We build on established additive architectures and fairness regularization and will make these connections more explicit. FNNC uses a conventional feedforward network and does not provide the additive score decomposition central to our analysis (Padala and Gujar, 2020, Section 4.1). Table 1 (p. 4) already compares regularized and unregularized FAN within the same model family. This addresses the effect of fairness regularization in our framework. Comparisons with separately implemented GAMI-Net or FNNC would provide additional benchmarking, but the current results are not intended to establish superiority over those alternatives.

  2. Historical lending decisions and the interpretation of fairness: The empirical target is stated in the paper. Section 4 (p. 4) identifies the outcome as denial versus origination and distinguishes this from default prediction. Section 8 (p. 9) states that assessing welfare, profitability, and fairness conditional on true borrower risk requires richer data. Our empirical finding concerns reduced group disparities in prediction errors relative to historical underwriting decisions, while preserving predictive fit. Appendix C (pp. 15–17) considers repayment outcomes in a separate theoretical setting.

  3. Proposition 2 and threshold-averaged EO disparity: Proposition 2 establishes a valid lower bound which cannot be inverted. Appendix F (p. 26) already acknowledges the limitation concerning pointwise EO, while the evidence that this disparity decreases is empirical: Appendix H, Table 8 (p. 35) directly reports a reduction in D1+D0D_1+D_0 from 0.00114 to 0.00046. Appendix I (pp. 37–40) separately evaluates fairness at operational thresholds. These results support the observed improvements without implying that minimizing the mean-score penalty universally guarantees them.

  4. Lemma 3 and learned subnetworks: Data-adaptive learning and component selection can be covered by a uniform bound over the complete, prespecified hypothesis class. However, the finite-dictionary argument in Appendix G.3 (pp. 32–33) does not complete that justification for FAN’s learned subnetworks. The asserted O(L/n)O(L/\sqrt n) rate also does not follow from bounded-domain Lipschitzness alone for the bivariate components. We will therefore formulate the generalization result using the Rademacher complexity of the full probability-function class, including all admissible subnetworks and component selections, and remove the unsupported simplification. This preserves the uniform-convergence approach while correcting the scope of the stated bound.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.