Abstract
Machine learning models often fail to generalize fairness across domains where sensitive attributes vary. This paper introduces SISA (Selective Invariance for Sensitive Attributes), a framework designed for Fair Domain Generalization (FDG) that handles heterogeneous sensitive attributes across different domains. Unlike traditional methods that assume uniform sensitive attributes, SISA employs a dual-representation architecture with selective invariance to maintain high predictive accuracy while enforcing fairness constraints specific to each domain's requirements.
The Challenge of Heterogeneous Fairness
Machine learning models frequently encounter "Domain Generalization" issues when deployed in target environments that differ from their training data. While ensuring fairness (e.g., preventing discrimination based on race or gender) is a priority, existing state-of-the-art methods typically assume that every domain shares the same sensitive attributes.
In real-world scenarios, legal and social priorities vary by region. A hiring AI in the US might focus on race and gender, while the same AI in another country might prioritize age or marital status. Current approaches would require training $2^n$ separate models to cover every combination of $n$ sensitive attributes, which is computationally prohibitive. SISA addresses this by creating a single model capable of generalizing both performance and fairness to unseen domains with varying sensitive requirements.
Methodology: Selective Invariance (SISA)
The SISA framework utilizes two distinct feature representations to balance performance and fairness:
- Predictive Representation ($z_g$): An encoder learns features that are invariant across all domains to ensure robust performance on unseen data.
- Fairness Representation ($z_f$): This representation is selectively invariant, meaning it is only forced to be similar across domains that share the same sensitive attributes.
To achieve this, the authors utilize Sensitivity Encoding, where inputs are paired with binary vectors indicating active sensitive attributes. They also employ a Domain Translator (such as StarGAN) during training to simulate data points across different contexts while preserving labels. A specialized Contrastive Loss ($L_{DF}$) aligns representations for domains sharing attributes while separating those that do not.
Experimental Evaluation
The researchers evaluated SISA on three diverse benchmarks:
- CelebA: A face dataset using "male" and "young" as sensitive variables.
- MIMIC-CXR: A medical imaging dataset (chest X-rays) focusing on Cardiomegaly, Edema, and Pneumothorax tasks with age, race, and gender as sensitive attributes.
- FACET: A computer vision benchmark with 32k images and 52 classes, designed for evaluating person-related attribute fairness.
Key Results and Insights
SISA consistently outperformed existing baselines such as FATDM and ERM:
- Accuracy Gains: On CelebA, SISA improved average AUROC by 1.46% over previous state-of-the-art methods. On the FACET dataset, it improved prediction accuracy by 5.2%.
- Superior Fairness: The model achieved near-zero unfairness scores (Mean Distance and Earth Mover’s Distance) on tasks where previous models remained biased.
- Efficiency: SISA provides these benefits using a single model, avoiding the need for the 3 to 15 separate models required by alternative approaches.
- Robustness to Correlations: The framework effectively separated correlated attributes (e.g., gender and cosmetic features) by adjusting the margin hyperparameter in its contrastive loss.
Conclusion and Future Work
SISA demonstrates that selective invariance is a more robust approach than total invariance, as it preserves more useful predictive information while meeting local fairness constraints. While the current model requires a pre-defined "universe" of sensitive attributes during training, future work aims to extend this to "zero-shot fairness," where the model can adapt to entirely new sensitive attributes it has never encountered before.
Summary updated