Technical Summary: Counterfactual Explanation Consistency for Procedural Fairness in Credit Decisions

Problem Statement

Current algorithmic fairness research in socially sensitive domains, such as credit lending, predominantly focuses on outcome fairness (e.g., demographic parity, equalized odds). These metrics ensure that predictive outcomes are statistically balanced across protected groups (e.g., race, gender). However, the authors argue that satisfying outcome fairness does not guarantee that the model utilizes the same reasoning process for individuals from different groups.

This gap creates a "hidden procedural bias" where a model may produce identical outcomes for comparable applicants from different demographic groups but rely on fundamentally different feature attributions to reach those conclusions. For instance, a model might approve a male applicant based primarily on credit score while approving a female applicant with a similar profile based on employment history. Under legal frameworks like the Equal Credit Opportunity Act (ECOA), this constitutes disparate treatment, yet it remains invisible to standard outcome-based fairness metrics.

The authors formalize this phenomenon using a 2×2 taxonomy of fairness regimes:

• Regime A (Fully Fair): Same prediction, same reasoning.
• Regime B (Hidden Bias): Same prediction, different reasoning (the critical blind spot).
• Regime C (Outcome Unfair): Different prediction, same reasoning.
• Regime D (Severely Unfair): Different prediction, different reasoning.

Existing methods primarily target Regimes C and D, leaving Regime B undetected.

Methodology: Counterfactual Explanation Consistency (CEC)

To detect and mitigate Regime B, the authors propose Counterfactual Explanation Consistency (CEC), a framework that aligns feature attributions between individuals and their counterfactual counterparts.

1. Counterfactual Generation via Label-Stratified Nearest-Neighbor Matching

Instead of using naive counterfactuals (flipping the protected attribute while holding all other features constant, which often yields statistically impossible individuals), the authors propose a Label-Stratified Nearest-Neighbor (NN) approach.

• Mechanism: For a factual individual $(x, y, a)$, the system searches the training data for an individual $\tilde{x}$ from the opposite group ($1-a$) who shares the same true label ($y$) and has the most similar financial features ($F$).
• Financial Features ($F$): Defined as merit-based, legally permissible features (e.g., income, credit score) that are not demographic proxies.
• Rationale: This ensures the factual and counterfactual have identical ground-truth creditworthiness. Any difference in the model's reasoning is thus attributable to group-dependent bias rather than legitimate risk differences. This method avoids the need for structural causal models (SCMs).

2. Consistent Baseline Principle

The framework utilizes Integrated Gradients (IG) to measure feature attributions. A critical contribution is the Consistent Baseline Principle.

• Problem: Comparing IG vectors using group-specific baselines (e.g., mean of Group 0 vs. mean of Group 1) introduces a confounder. Differences in attributions may arise simply from the different reference points (baselines) rather than discriminatory reasoning.
• Solution: Both the factual $x$ and counterfactual $\tilde{x}$ must be evaluated against the same baseline (specifically, the mean of the factual's label-group).
• Theoretical Guarantee: The authors prove that using a consistent baseline isolates the attribution difference to the input difference and the model's gradients, removing spurious terms caused by group distribution shifts.

3. The CEC Metric and Training Loss

• Metric: The CEC score measures the normalized $L_2$ distance between the attribution vectors of $x$ and $\tilde{x}$ when using the consistent baseline. A score of 0 indicates identical reasoning; 1 indicates maximally opposed reasoning.
• Training Objective: The authors introduce a multi-objective loss function that jointly optimizes:
  1. Prediction Accuracy (Binary Cross-Entropy).
  2. Outcome Fairness (Equalized Odds constraint).
  3. Procedural Fairness (CEC Loss): A differentiable regularizer that penalizes divergence in attribution vectors between factual and counterfactual pairs.

Key Contributions

1. Formalization of Hidden Procedural Bias: The paper defines and categorizes "Regime B" (same outcome, different reasoning) as a critical failure mode invisible to standard fairness metrics.
2. Nearest-Neighbor Counterfactual Generation: A practical method to generate realistic counterfactuals for fairness auditing without requiring causal graphs, relying instead on matching financial merit.
3. Consistent Baseline Principle: A theoretical and practical correction for Integrated Gradients comparisons that prevents baseline-induced artifacts from being misinterpreted as bias.
4. CEC Framework: A differentiable training loss that enforces explanation consistency alongside accuracy and outcome fairness, demonstrating that procedural fairness can be achieved with modest utility costs.

Experimental Results

The authors evaluated CEC on four datasets: Synthetic data, German Credit, Adult Income, and HMDA mortgage data. They compared CEC against six baselines (Unconstrained, Disparate Impact Remover, Hardt Post-Processing, Agarwal Reductions, Adversarial Debiasing, Lagrangian Fair Learning).

• Detection of Hidden Bias: All baseline methods that achieved competitive outcome fairness (low Equalized Odds gaps) exhibited substantial hidden procedural bias (high CEC scores). For example, on the German Credit dataset, baseline methods placed ~90% of samples in Regime B, whereas the CEC model reduced this to ~17%.
• Mitigation: The CEC-trained model consistently achieved the lowest CEC scores and the highest proportion of Regime A (fully fair) samples across all datasets.
• Utility Trade-off: The CEC model maintained competitive predictive utility (F1 and AUC) and outcome fairness, with only modest reductions in performance compared to unconstrained baselines.
• Ablation Studies:
  • Adding Equalized Odds (EO) constraints alone did not reduce Regime B; in some cases, it slightly increased it.
  • Adding CEC regularization alone significantly reduced Regime B and, as a side effect, often drove the EO gap to zero.
  • The full multi-objective model achieved the best balance, confirming that procedural fairness and outcome fairness are orthogonal objectives that require explicit joint optimization.

Significance and Claims

The paper claims that outcome fairness and procedural fairness are orthogonal. Optimizing for one does not guarantee the other; in fact, standard outcome-focused interventions can inadvertently mask or exacerbate hidden procedural bias (Regime B).

The significance of this work lies in:

1. Regulatory Compliance: It addresses a specific gap in legal compliance (e.g., ECOA), where lenders are prohibited from using different underwriting standards for different groups, even if the final approval rates are equal.
2. Auditability: It provides a quantitative tool (CEC score and regime distribution) to audit not just what a model decides, but how it reasons, identifying specific applicants subjected to inconsistent treatment.
3. Methodological Shift: It demonstrates that procedural fairness can be integrated into the training pipeline via differentiable losses, moving beyond post-hoc explanations to proactive mitigation of hidden bias.

The authors conclude that while outcome fairness is necessary, it is insufficient for ensuring equitable automated decision-making in high-stakes domains like lending. The CEC framework offers a practical path to align model reasoning with legal and ethical standards of procedural justice.