A Fairness-Aware and Bias-Resilient XAI Framework for Equitable Financial Decision-Making

Artificial intelligence increasingly influences financial decisions, including loan approvals and credit scoring, yet historical and structural biases embedded in data and processes can propagate through predictive models, creating disparate outcomes across demographic groups [1-5]. Regulatory and ethical expectations further require decisions to be transparent and explainable, with reasoned, auditable narratives for adverse actions and with safeguards that balance fairness and predictive utility [5-7]. The task considered is loan‑approval prediction with feature set X, protected attribute A (analyzed one at a time: race, ethnicity, sex, or age), and target Y (approval vs. denial). Fairness is assessed using Demographic Parity (DP), Equal Opportunity (EO), and Equalized Odds (EOdds).

Single‑stage debiasing and post‑hoc fixes are insufficient to ensure durable equity in operational settings because bias can re‑enter through data shifts, proxy features, or inconsistent thresholds; deployment‑time monitoring and governance are therefore required to maintain fairness without eroding predictive performance [2, 8, 9]. Beyond centralized modeling, privacy, regulatory, and organizational constraints motivate training regimes where data remain local and only model updates are aggregated, introducing additional fairness considerations under non‑IID distributions [10, 11]. These needs call for end‑to‑end frameworks that integrate mitigation across the model lifecycle and connect technical interventions to compliance‑oriented explanations and controls.

Fairness‑Aware, Interpretable, Resilient, and Equitable (FAIRE) addresses these challenges as a multi‑stage pipeline. At the data level, representation is balanced via reweighing and sampling; at the in‑training level, adversarial debiasing and fairness‑regularized objectives reduce dependence on sensitive attributes; at the post‑processing level, calibrated, group‑specific thresholds and reject‑option rules align decisions with policy constraints. Transparency is supported by global and local explainability—permutation importance, interpretable surrogates, LIME, SHAP, and Integrated Gradients to generate reason codes suitable for adverse‑action notices [6]. To accommodate organizational data boundaries, a federated learning (FL) regime enables client‑local training with privacy-preserving aggregation while auditing fairness at each round [6, 10].

The contributions of the study are listed below:

• End-to-end framework with numeric fairness reporting using DP, EO, and EOdds.
• Federated evaluation under non-IID client partitions with per-round support‑weighted fairness aggregation [10].
• Composite Interpretability Score that synthesizes surrogate fidelity, explanation sparsity, and stability, grounded in model‑agnostic and model-specific Explainable AI (XAI) [6].
• Reproducible baselines and statistical testing aligned to prior art, enabling credible comparisons under clearly specified protocols [5, 8, 11].

Let the dataset be $D=\left\{\left(X_i, A_i, Y_i\right)\right\}_{i=1}^N$, where $X$ denotes predictive features, $A$ denotes a single protected attribute considered per analysis (race, ethnicity, sex, or age), and $Y$ denotes the loan-approval outcome (originated vs. denied), as defined in Eq. (1).

Two training regimes are employed (detailed in Section 3.1). The centralized track trains a single model on pooled data with fairness‑aware objectives and post‑processing as specified in Section 3.3. The federated track distributes training across clients that update local models and periodically aggregate parameters; both regimes feed the same bias‑detection, explainability, and monitoring components described in Sections 3.2–3.5.

Fairness is evaluated using DP, EO, and EOdds, defined in Eqs. (5)-(7). These metrics are computed pre‑training to characterize historical disparities, during training to guide mitigation, and post‑training to quantify residual bias; corresponding results are reported in Section 4 alongside classification and interpretability outcomes.

3.1 Architecture overview

Figure 1 shows the FAIRE pipeline as an integrated sequence of modules: data ingestion and preprocessing, bias‑mitigation layers (pre‑, in‑, and post‑training), explainability, deployment and decisioning, and monitoring and governance. During ingestion, sensitive attributes A are separated from predictive features X in accordance with Eq. (1), and feature transformation procedures reduce proxy leakage while preserving predictive signal. Typical transformations include coarsening or binning highly identifying variables, constrained encodings, and leakage checks to ensure that A does not re‑enter the modeling stack through correlated surrogates.

$D=\left\{\left(X_i, A_i, Y_i\right)\right\}_{i=1}^N$                        (1)

Figure 1. FAIRE framework pipeline

Centralized track: A single predictor $f_\theta$ is trained end-to-end with an adversary $g_\phi$ that attempts to infer $A$ from the learned representation. The joint objective in formula (2) balances prediction loss against the adversary's loss via the trade-off parameter $\lambda$, encouraging $f_\theta$ to discard information predictive of $A$ while retaining signal for $Y$. Data-level mitigation (reweighing/oversampling) may be applied before training, and a fairness regularizer may be included in the loss as specified later in Section 3.3. After training, group-specific operating thresholds and a reject-option band can be applied to the centralized scores following Eqs. (12) and (13).

Federated track: Data remain local to each client k. Each client trains a local predictor $f_{\theta_k}$ with a local adversary $g_{\phi_k}$ under the same objective as formula (2), using local data shards and identical preprocessing/mitigation rules. Model updates are transmitted to a coordinator that aggregates client parameters using FedAvg; unless otherwise specified, the protocol uses T = 100 communication rounds with E = 2 local epochs per round. Only model updates or parameters are shared—no raw records, explanations, or identifiers. After each aggregation, the coordinator evaluates global fairness diagnostics using privacy‑preserving summaries and, when needed, recommends calibrated group‑specific thresholds consistent with Eqs. (12) and (13). This two‑track design enables consistent fairness controls irrespective of training regime while preserving locality of data.

$\min _\theta \mathbb{E}\left[\mathcal{L}\left(f_\theta(X), Y\right)\right]-\lambda \cdot \mathbb{E}\left[\mathcal{L}_{a d v}\left(g_\phi\left(f_\theta(X)\right), A\right)\right]$                  (2)

$\phi_j=\sum_{S \subseteq\{1, \ldots, d\} \backslash\{j\}} \frac{|S|!(d-|S|-1)!}{d!}[f(S \cup\{j\})-f(S)]$                  (3)

Explainability layer: Explanations are produced at global and local levels for either track. Global diagnostics include permutation importance and surrogate modeling (Eqs. (14) and (15), supporting system‑level audits of feature influence and fidelity. Local diagnostics include LIME, SHAP, and Integrated Gradients (Eqs. (16)-(18)), yielding per‑applicant reason codes and ranked contributions suitable for compliance review and adverse‑action communication. The explainability interface is identical across tracks and attaches to the scoring output, enabling uniform monitoring and reporting.

Deployment and decisioning: The scoring service outputs calibrated approval probabilities with accompanying explanation artifacts. Group‑specific thresholds and reject‑option rules are implemented as decision‑policy layers on top of the raw scores, following Eqs. (12) and (13). Policy parameters are versioned and auditable, allowing subsequent monitoring to attribute fairness movements to either model updates or threshold adjustments.

Monitoring: Fairness and performance drift are tracked during validation and post‑deployment. The trigger in formula (4) detects stepwise changes in DP across evaluation windows, summarizes alert states, recommended actions, and intervention history. Selection of the drift tolerance $\delta$, alert routing, and human‑in‑the‑loop review are specified in Section 3.5, ensuring consistent governance across centralized and federated operating modes.

$\left|D P_t-D P_{t-1}\right|>\delta$                          (4)

Data ingestion separates $A$ from $X$ per Eq. (1); mitigation spans pre-training reweighing/oversampling, in-training adversarial debiasing (formula (2)), and post-processing thresholds (Eqs. (12) and (13)); explainability and deployment produce decisions with reason codes; monitoring applies drift checks (formula (4)). The federated branch depicts client $k$ local training with adversary $g_{\phi_k}$, rounds $t=1 \ldots 100$ with local epochs $E=2$, and server-side aggregation $\mathcal{A}=$ FedAvg.

3.2 Bias detection

Metrics: DP, EO, and EOdds are adopted as in Eqs. (5)-(7). DP assesses parity of approval rates across groups in Eq. (5):

$D P=\frac{P(\hat{Y}=1 \mid A=0)}{P(\hat{Y}=1 \mid A=1)}$                        (5)

EO measures parity of true‑positive rates among qualified applicants in Eq. (6):

$\begin{gathered}E O=P(\hat{Y}=1 \mid Y=1, A=0)-P(\hat{Y}=1 \mid Y=1, A=1)\end{gathered}$                        (6)

EOdds aggregates disparities in both true‑positive and false‑positive behavior in Eq. (7):

$\begin{aligned} \text { EOdds } & =\mid P(\hat{Y}=1 \mid Y=1, A=0)  -P(\hat{Y}=1 \mid Y=1, A=1) \mid  +\mid P(\hat{Y}=1 \mid Y=0, A=0)-P(\hat{Y}  =1 \mid Y=0, A=1) \mid\end{aligned}$                        (7)

Federated fairness metrics: In the federated track, each client k computes group‑wise confusion matrices and derives local approval rates and error rates. Only privacy‑preserving summaries are shared: per‑group supports and rate estimates (e.g., approval $\mathrm{s}_{k, a}=\mathrm{TP}_{k, a}+\mathrm{FP}_{k, a}$, and total $\mathrm{s}_{k, a}=P_{k, a}+ \left.N_{k, a}\right)$). Server‑side aggregation forms global metrics from support‑weighted counts rather than averaging ratios. For example:

$\begin{aligned} & \bullet \mathrm{TP}_{k, a}, \mathrm{FP}_{k, a}, \mathrm{TN}_{k, a}, \mathrm{FN}_{k, a} . \\ & \bullet \operatorname{Pr}(\hat{Y}=1 A=a)=\frac{\sum_{k=1}^K \text { approvals }_{k, a}}{\sum_{k=1}^K \text { totals }_{k, a}}, \\ & \bullet \mathrm{TPR}_a=\frac{\sum_{k=1}^K \mathrm{TP}_{k, a}}{\sum_{k=1}^K \mathrm{P}_{k, a}}, \mathrm{FPR}_a=\frac{\sum_{k=1}^K \mathrm{FP}_{k, a}}{\sum_{k=1}^K \mathrm{~N}_{k, a}},\end{aligned}$

where, TPR refers to True Positive Rate, and FPR indicates False Positive Rate.

Which are then combined into DP, EO, and EOdds using the definitions above. Logging stores only aggregated counts and rates to maintain privacy.

Evaluation process: Metrics are computed pre‑training to characterize baseline disparities, during training per epoch (centralized) or per round (federated) to guide mitigation, and post‑training to quantify residual bias. After deployment, the same metrics are monitored on rolling windows; drift detection follows Eq. (4) and thresholding logic in Section 3.5, with alerts routed to governance for review and action.

3.3 Bias mitigation techniques

Data‑level (pre‑training): Reweighing adjusts the empirical loss by assigning each instance a weight based on the joint distribution of the protected attribute and the label, as in Eq. (8). Define the group‑label weight $w_{a, y}=\frac{\operatorname{Pr}(A=a) \operatorname{Pr}(Y=y)}{\operatorname{Pr}(A=a, Y=y)}$ and minimize a weighted loss in Eq. (8):

$\mathcal{L}=\sum \mathrm{w}_{\mathrm{i}} \cdot \mathcal{L}\left(\mathrm{f}\left(\mathrm{X}_{\mathrm{i}}\right), \mathrm{Y}_{\mathrm{i}}\right)$                   (8)

Oversampling complements reweighing by increasing minority‑group support until approximate parity is achieved, targeting as in formula (9):

$n_0 \approx n_1$                   (9)

Synthetic sampling may be applied when raw counts remain insufficient, with guardrails: generate only within the convex hull of observed minority‑group feature vectors; preserve label‑conditional distributions; prohibit direct use of A or near‑deterministic proxies during synthesis; perform synthesis on training folds only; and validate that downstream calibration and ranking are not distorted by synthetic artifacts.

Model‑level (in‑training): Adversarial debiasing follows formula (10):

$\min _\theta \mathbb{E}\left[\mathscr{L}\left(f_\theta(X), Y\right)\right]-\lambda \cdot \mathbb{E}\left[\mathscr{L}_{a d v}\left(g_\phi\left(f_\theta(X)\right), A\right)\right]$                   (10)

Optimizing a predictor $f \theta$ for the approval task while an adversary $g \theta$ attempts to infer A from intermediate representations. The training objective balances prediction loss and adversarial loss via a trade‑off parameter λ. A fairness regularizer augments the objective as in Eq. (11):

$\mathscr{L}=\mathbb{E}[\mathscr{L}(f(X), Y)]+\alpha R$                   (11)

where, α = 0.01. The adversary $g_\phi$ is a 2‑layer MLP (64–32) with ReLU activations, gradient‑reversal for signal inversion, and dropout 0.1. A λ‑sweep {0, 0.1, 0.2, …, 1.0} traces the accuracy–fairness frontier and supplies operating points later selected under explicit constraints. Centralized optimization uses batch size 1024, learning rate 1 × 10⁻³, and early‑stopping patience 5; the federated track mirrors these settings locally and aggregates with FedAvg for T = 100 rounds with E = 2 local epochs per round. These choices keep optimization stable while exposing a broad fairness‑regularization envelope suitable for post‑hoc selection under policy constraints.

Post‑processing: Decision‑policy adjustments refine deployed behavior after score production. Group‑specific thresholds implement Eq. (12):

$T_0=T_1+\Delta T$                   (12)

where, ΔT is tuned on validation data to minimize the EO gap subject to an AUC decrease ≤ 0.01 (one percentage point). When borderline uncertainty dominates disparities, a reject‑option band per Eq. (13) introduces a small margin γ = 0.02 around the decision boundary: classify and defer or route to manual review otherwise.

$\widehat{Y}=1, \quad$ if $P(Y=1 \mid X) \geq T+\gamma$                   (13)

Thresholds and margins are versioned for auditability and are applied consistently across centralized and federated tracks.

Hyper‑parameters and schedule (centralized and federated): Unless stated otherwise, training uses batch size 1024, learning rate 1 × 10⁻³, and early‑stopping patience 5. The federated protocol applies FedAvg with T = 100 communication rounds and E = 2 local epochs per round; clients adopt the same optimizer and batch size as the centralized track. These settings interact with the λ‑grid and α to produce a family of models spanning differing fairness–utility trade‑offs, enabling principled operating‑point selection under the evaluation protocol.

3.4 Explainable AI component

Global interpretability: Global analysis combines permutation feature importance and an interpretable surrogate. Permutation importance measures the loss increase when a single feature is randomly shuffled, yielding an importance score $I\left(x_j\right)$ as formalized in Eq. (14):

$I\left(X_j\right)=\mathbb{E}\left[\mathscr{L}(f(X))-\mathscr{L}\left(f\left(X_{-j}\right)\right)\right]$                   (14)

A surrogate model g is then trained to approximate the scoring function f by minimizing the discrepancy between $g(X)$ and $f(X)$ as in formula (15):

$\min \sum\left(f\left(X_i\right)-g\left(X_i\right)\right)^2$                   (15)

Fidelity is summarized by R² or an equivalent bounded error metric. Together, these tools characterize system‑level reliance on features and provide a stable global view that complements fairness diagnostics.

Local interpretability: Instance‑level explanations are produced using three complementary methods. LIME fits a simple, locally weighted model g around a perturbed neighborhood of the instance, optimizing the locality‑aware objective:

$\phi_j=\sum_{S \subseteq\{1, \ldots, d\} \backslash\{j\}} \frac{|S|!(d-|S|-1)!}{d!}[f(S \cup\{j\})-f(S)]$                   (16)

This offers piecewise‑linear insight near the decision point. SHAP attributes a Shapley value $\phi_j$ to each feature using the cooperative‑game formulation in Eq. (17):

$\phi_j=\sum_{S \subseteq\{1, \ldots, d\} \backslash\{j\}} \frac{|S|!(d-|S|-1)!}{d!}[f(S \cup\{j\})-f(S)]$                   (17)

The method provides additivity and consistency, enabling faithful “reason codes.” Integrated Gradients accumulates path‑integrated gradients from a baseline x′ to the instance x as in Eq. (18):

$I G_j(x)=\left(x_j-x_j^{\prime}\right) \int_{\alpha=0}^1 \frac{\partial f\left(x^{\prime}+\alpha\left(x-x^{\prime}\right)\right)}{\partial x_j} d \alpha$                   (18)

This is preferred for differentiable deep models where gradient information is available. Operational guidance: LIME for quick, human‑readable local approximations; SHAP for axiomatic attributions and ranked reasons; Integrated Gradients for deep networks requiring gradient‑path attributions.

Composite Interpretability Score: A bounded, composite score aggregates complementary qualities of explanations, tied to Eqs. (14)-(18): IS $=w_1 \cdot \operatorname{Fidelity}(g, f)+w_2 \cdot$ Sparsity $_\tau+w_3$. Stability ${ }_{k, B}$, with $w_1=0.50, w_2=0.25$, $w_3=0.25$; threshold $\tau=0.01$ on $\left|\phi_j(x)\right|$ (SHAP magnitude); top $-k=5$; bootstraps $B=1000$. Components are computed as follows:

$\begin{aligned} & \bullet \text {Fidelity}(g, f) \in[0,1] \text {,e.g., } R^2=-\frac{\sum_i\left(f\left(X_i\right)-g\left(X_i\right)\right)^2}{\sum_i\left(f\left(X_i\right)-f\right)^2} \\ & \bullet \text {Sparsity}_\tau=1-\mathbb{E}_x\left[\frac{1}{d} \sum_{j=1}^d 1\left\{\left|\phi_j(x)\right|>\tau\right\}\right] \in[0,1] \\ & \bullet \text {Stability}_{k, B}=\mathbb{E}_x\left[\frac{2}{B(B-1)} \sum_{b<b^{\prime}} \frac{\left|T_k^{(b)}(x) \cap T_k^{\left(b^{\prime}\right)}(x)\right|}{\left|T_k^{(b)}(x) \cup T_k^{\left(b^{\prime}\right)}(x)\right|}\right] \in[0,1]\end{aligned}$

where, $T_k^{(b)}(x)$ is the set of top $-k$ features by $\left|\phi_j(x)\right|$ under bootstrap $b$. The Interpretability Score increases with surrogate faithfulness, succinct reason sets, and bootstrap-stable attributions.

Regulatory explanations: Adverse‑action narratives combine SHAP‑ranked drivers with threshold logic formalized as below. A de‑identified template is:

$I<T \Rightarrow$ Application Denied(Insufficient Income)                   (19)

3.5 Monitoring and governance

Automated scanner and cadence: Fairness metrics—DP, EO, and EOdds—are recomputed on a weekly cadence using rolling windows. Evaluations occur at two granularities: (i) global aggregates over the full scoring stream and (ii) per‑client aggregates in the federated regime to surface localized drift. For each evaluation window, prediction outputs, per‑group confusion matrices, and derived rates are generated; corresponding SHAP attribution vectors and summary statistics are persisted to an immutable artifact store alongside model version, data slice identifiers, decision thresholds, and federated aggregation metadata. This schedule aligns post‑deployment monitoring with the training‑time fairness checks and supports traceability to formula (4) and the operational monitoring formulation (20):

$\left|D P_t-D P_{t-1}\right|>\delta$                   (20)

Drift thresholds and decision rules: The drift rule in formula (20) is retained with δ = 0.05 for demographic‑parity movements. DP values within [0.80,1.25] are treated as in‑band (80% rule). For each window and group, bootstrap confidence intervals (CIs) are computed for DP, EO, and EOdds using the stored counts; an alert is triggered when the CI for DP lies entirely outside [0.80,1.25] or when EO/EOdds CIs exclude zero. Alert payloads include: metric name, affected group(s), point estimates with CIs, window identifiers, current thresholds, and suggested actions (ΔT adjustment per Eq. (12), reject‑option margin tuning per Eq. (13), or retraining). All computations rely on privacy‑preserving aggregates produced by the scoring service and federated clients.

Alert routing and service levels: Alerts are routed programmatically to model risk management and compliance queues. An escalation service‑level agreement of 5 business days governs review and sign‑off. Each alert results in a case record with: evidence artifacts (metrics, CIs, SHAP summaries), proposed remediation, reviewer determination, and effective‑date timestamps. Overrides and deferrals are logged with rationale and risk classification; follow‑up evaluations verify remediation efficacy at the next scheduled window. Human review and policy alignment are integral steps of the governance loop described for deployed systems.

Runbook of remediations: When alerts are confirmed, actions follow a minimal‑impact sequence: (1) threshold recalibration via ΔT to restore DP/EO toward target bands while constraining AUC change; (2) reject‑option activation with a small margin γ to reduce boundary‑region disparities; (3) scheduled retraining with the selected λ operating point from Section 3.3 and refreshed data; (4) federated‑specific measures, such as client‑level thresholding or reweighting, when drift localizes to a subset of clients. Post‑action monitoring verifies resolution using the same CI‑based decision rules.

Dashboard and reporting: The monitoring dashboard presents three synchronized panels backed by the audit store:

• Panel A: approval‑rate trends by protected group with a shaded DP target band [0.80,1.25].
• Panel B: drift alerts derived from formula (20), visualizing δ = 0.05 bands, recent violations, and time‑to‑resolution.
• Panel C: an action log enumerating threshold changes (ΔT), reject‑option activations, retraining events, and reviewer outcomes.

Artifacts refresh weekly; data sources are scoring outputs (predictions, per‑group counts) and fairness audit logs (metrics, CIs, actions).

Panel A: approval‑rate trends by protected group with DP target band [0.80,1.25]. Panel B: fairness‑drift alerts from formula (20) with δ = 0.05 bands and current alert states. Panel C: action log summarizing applied mitigations (ΔT thresholds, reject‑option), retraining events, and reviewer dispositions; weekly refresh; data from model‑scoring outputs and fairness audit logs.

At the selected operating point, the study demonstrates measurable improvements in group equity with minimal utility cost. DP rises from 0.74 [0.72,0.76] to 0.92 [0.90,0.94], the EO (TPR) gap declines from 0.16 [0.15,0.17] to 0.05 [0.04,0.06], and Equalized Odds decreases from 0.23 [0.21,0.25] to 0.07 [0.06,0.09]. AUC‑ROC remains high at 0.88 [0.87,0.89], within 0.5 percentage points of the high‑utility setting. The fairness–utility frontier indicates feasible operating regions that satisfy a demographic‑parity target band of [0.80,1.25] without material loss of predictive accuracy. Federated training with 50 clients achieves performance and fairness close to centralized post‑mitigation while accommodating non‑IID client distributions. Final global outcomes reach AUC‑ROC 0.87 [0.86,0.88], DP 0.90 [0.88,0.92], EO 0.06 [0.05,0.07], and EOdds 0.11 [0.10,0.12]; confidence intervals widen for clients with limited protected‑group support. Support‑weighted aggregation of client summaries mitigates small‑sample volatility and yields stable global metrics, consistent with observations in federated‑fairness and dataset‑bias surveys. Interpretability outcomes indicate improvements in a composite Interpretability Score through higher surrogate fidelity, sparser reason sets, and more stable attributions. Global drivers—such as debt‑to‑income ratio, credit score, and loan‑to‑value—align with domain expectations, and threshold‑style explanations supply adverse‑action reasons derived from the leading local contributors. Continuous fairness monitoring with a fixed drift tolerance of 0.05 maintains in‑band behavior over rolling evaluation windows, supporting governance and auditability. Limitations include potential proxy leakage and missingness in key variables, the focus on single‑attribute fairness rather than intersections, temporal and lender heterogeneity that may affect transportability, and communication/compute overheads in federated settings. Future work includes adaptive thresholding and λ‑selection under explicit constraints, integration of differential privacy in federated pipelines, intersectional and causal analyses of disparities, automated remediation policies triggered by monitoring alerts, and longitudinal deployments with regulatory audits. To sum up, FAIRE achieves in‑band fairness with minimal utility change across centralized and federated regimes while producing audit‑ready explanations and sustaining an operational monitoring loop suitable for real‑world credit decisioning.