Procedural Fairness via Group Counterfactual Explanation

1. Problem Statement

Current fairness research in machine learning predominantly focuses on outcome-oriented fairness (e.g., Equalized Odds, Statistical Parity), which ensures that model predictions do not systematically disadvantage protected groups. However, these metrics only evaluate what decisions are made, not how they are reached.

The paper identifies a critical gap: two models can achieve identical outcome fairness metrics while relying on fundamentally different decision-making processes for different groups. This lack of procedural fairness means a model might use different reasoning logic for Group A versus Group B, even for individuals with the same true label. Existing explanation methods are typically post-hoc (diagnostic), used to audit models after training, rather than being integrated into the learning process to prevent disparate reasoning from occurring in the first place.

Core Question: How can we enforce that a model uses consistent reasoning (explanations) across protected groups, conditional on the true label, without sacrificing predictive performance?

2. Methodology: Group Counterfactual Integrated Gradients (GCIG)

The authors propose GCIG, an in-processing regularization framework that enforces explanation invariance. The core idea is to penalize the model if its explanation for a specific prediction changes significantly when the "context" (the reference group) changes, even if the input and true label remain the same.

Key Components:

• Group Conditional Baselines: Instead of using a single global baseline (e.g., a zero vector), GCIG computes a baseline for each protected group $g$ and label $y$.
  $$b_{y,g} = \mathbb{E}[X | Y=y, A=g]$$
  This represents the average feature profile of individuals in group $g$ who share the same true outcome $y$. These baselines are updated using Exponential Moving Averages (EMA) during training to ensure stability.
• Group Counterfactual Integrated Gradients (IG): For a given input $x$ with true label $y$, the model computes Integrated Gradients relative to each group's baseline ($b_{y,0}$ and $b_{y,1}$).
  $$IG^{(g)}(x; y) = IG(x; b_{y,g})$$
  This answers the counterfactual question: "How does the model explain this prediction if we assume the individual belongs to group $g$'s typical context?"
• Explanation Disparity Metric: The method normalizes these attribution vectors to remove scale effects and calculates the variance of these attributions across groups.
  $$V(x; y) = \text{Var}_{g \in A} \left[ \frac{IG^{(g)}(x; y)}{\|IG^{(g)}(x; y)\|} \right]$$
  A high $V(x; y)$ indicates the model relies on different features to justify the same outcome depending on the group context (procedural bias).
• Training Objective (FairX): The total loss function combines three components:
  $$L_{total} = L_{pred} + \lambda_{ig} L_{GCIG} + \lambda_{fair} L_{fair}$$
  • $L_{pred}$: Standard prediction loss (e.g., Cross-Entropy).
  • $L_{GCIG}$: The penalty for cross-group explanation variance.
  • $L_{fair}$: Standard outcome fairness constraint (e.g., Equalized Odds).
  • $\lambda$ terms control the trade-off between utility, outcome fairness, and procedural fairness.

3. Key Contributions

1. Formalization of Procedural Fairness: The authors define procedural fairness as Group Counterfactual Explanation Invariance. This requires that model explanations remain stable across protected groups when conditioned on the true label.
2. GCIG Framework: They introduce a novel in-processing regularization technique that integrates explanation constraints directly into the training loop, moving beyond post-hoc auditing.
3. Empirical Validation: The paper demonstrates that GCIG significantly reduces explanation disparity across multiple datasets while maintaining competitive predictive accuracy and outcome fairness.

4. Experimental Results

The authors evaluated FairX (the training framework using GCIG) on four benchmark datasets: Adult Income, German Credit, COMPAS, and Bank Marketing. They compared it against unconstrained models and state-of-the-art fairness baselines (Hardt's post-processing, Agrawal's Reductions, Adversarial training, etc.).

• Procedural Fairness (GCIG Score): FairX consistently achieved the lowest explanation disparity across all datasets.
  • Example: On the German Credit dataset, GCIG dropped from 0.190 (unconstrained) to 0.066.
  • Example: On COMPAS, it dropped from 0.193 to 0.034.
  • FairX won the "champion" status for procedural fairness on 100% of the datasets (4/4).
• Outcome Fairness & Utility:
  • FairX achieved the best Equalized Odds (EO) gap on German Credit and remained competitive on others.
  • Predictive performance (F1 score) was maintained or slightly improved compared to baselines. For instance, on German Credit, FairX achieved the highest F1 (0.833).
• Correlation Analysis: The study found a weak correlation ($r \approx 0.24$) between outcome fairness (EO) and procedural fairness (GCIG). This confirms that optimizing for equal error rates does not guarantee consistent reasoning; the two objectives are largely independent.
• Ablation Study:
  • Adding only outcome fairness (EO) actually worsened procedural fairness by 14%.
  • Adding only procedural fairness (GCIG) improved it by 6%.
  • Joint optimization (Full Model) yielded the best result, improving procedural fairness by 24% without degrading utility or outcome fairness.

5. Significance and Conclusion

This work is significant because it shifts the paradigm of fairness from purely outcome-based to process-based.

• Beyond Post-Hoc: It demonstrates that procedural fairness can be enforced during training, preventing disparate reasoning patterns from forming rather than just detecting them later.
• Complementarity: The results show that outcome fairness and procedural fairness are not mutually exclusive; in fact, they are complementary. Optimizing for both yields a more robust and trustworthy model.
• Trust and Transparency: By ensuring that the model's "reasoning" is invariant to group membership, the approach builds greater trust in AI systems, particularly in high-stakes domains like lending and criminal justice.

Limitations & Future Work: The current implementation focuses on binary protected attributes and tabular data. Future work aims to extend GCIG to multi-valued/intersectional attributes and other data modalities (text, images), as well as exploring the theoretical bounds of explanation invariance.