Abstract
Owing to the increasing use of machine learning in our daily lives, the problem of fairness has recently become an important topic in machine learning societies. Recent studies regarding fairness in machine learning have been conducted to attempt to ensure statistical independence between individual model predictions and designated sensitive attributes. However, in reality, cases exist in which the sensitive variables of data used for learning models differ from the data upon which the model is applied. In this paper, we investigate a methodology for developing a fair classification model for data with limited or no labels, by transferring knowledge from another data domain where information is fully available. This is done by controlling the Wasserstein distances between relevant distributions. Subsequently, we obtain a fair model that could be successfully applied to two datasets with different sensitive attributes. We present theoretical results validating that our approach provably transfers both classification performance and fairness over domains. Experimental results show that our method does indeed promote fairness for the target domain, while retaining reasonable classification accuracy, and that it often outperforms comparative models in terms of joint fairness.
Introduction
Machine learning models often inherit historical biases from their training data, leading to unfair treatment of certain groups based on "sensitive attributes" like race, age, or gender. While researchers have developed many techniques to make models "fair," these methods typically assume that the sensitive attribute remains constant between the training (source) and deployment (target) environments.
In real-world scenarios, this assumption often fails. A model trained on a dataset where race is the primary sensitive concern might be deployed in a region where gender is the critical factor. Furthermore, the target domain often lacks the labels needed to re-train the model for fairness. This paper introduces a novel framework for "joint transfer," allowing a model to carry over both its predictive accuracy and its fairness constraints across domains, even when sensitive attributes change.
Methodology
The proposed deep learning architecture uses an encoder to map input data into a latent space. The core of the approach is a joint loss function that balances three primary objectives:
- Accuracy: Standard supervised learning using cross-entropy loss on labeled source data.
- Source Fairness: Minimizing the 1-Wasserstein distance between the predicted score distributions of different groups in the source domain to ensure demographic parity.
- Domain Alignment: Using Sliced Wasserstein Discrepancy (SWD) to align the latent representations of corresponding groups across domains (e.g., mapping the "minority" group in the source to the "minority" group in the target).
Wasserstein distance (or "Earth Mover's Distance") is utilized because it considers the underlying geometry of the data, providing a meaningful measure of distributional distance even without overlap. The "Sliced" version (SWD) ensures computational efficiency by projecting high-dimensional data into one-dimensional slices.
Experimental Setup
The method was evaluated on three standard benchmarks for algorithmic fairness:
- UCI Adult: Census data used to predict income levels (Sensitive attributes: Race and Gender).
- ProPublica COMPAS: Criminal recidivism prediction (Sensitive attributes: Race and Gender).
- UCI German Credit: Credit risk assessment (Sensitive attributes: Age and Gender).
Experiments involved switching the roles of source and target sensitive attributes (e.g., training on Race and testing on Gender) to test the robustness of fairness transfer.
Key Results and Discussion
The proposed method consistently outperformed baselines in achieving "joint fairness." On the Adult dataset, transferring from Race to Gender, the method reduced the Demographic Parity gap to 0.0398, a significant improvement over standard domain adaptation (0.0781) and unfair models (0.2098).
As shown in the figures above, the model maintains similar positive prediction rates ($P(\hat{Y}=1|A)$) across different demographic groups in both the source and target domains. This indicates that fairness is transferable; by aligning the latent structures of demographic groups, the model learns a representation that is inherently "blind" to the disparities of both domains.
While there is a known trade-off between fairness and accuracy, the authors' method maintained high accuracy (often within 1-3% of the unfair baseline) while significantly improving fairness metrics. The authors also provided a mathematical "generalization bound" proving that controlling the Wasserstein distance between domains theoretically limits performance and fairness degradation.
Conclusion
This research demonstrates that fairness can be treated as a piece of transferable knowledge. However, the approach assumes some underlying similarity between domains; extreme differences could lead to "negative transfer." While achieving "Equal Opportunity" still requires limited target labels (approx. 20%), the framework provides a robust foundation for fair domain adaptation. Future work will explore handling multiple simultaneous sensitive attributes and individual fairness constraints.
Summary updated