Most algorithmic fairness techniques rely on a luxury that real-world practitioners rarely have: a dataset where every individual is labeled with sensitive attributes like race, gender, or age. In practice, privacy regulations, legal restrictions, or incomplete data collection often leave target datasets "blind" to these protected categories.
Recent research has converged on a powerful solution: Fairness Transfer. The core intuition is that while we may lack sensitive labels in our specific target domain, we can "borrow" the understanding of bias from a labeled source domain and project that knowledge onto our unlabeled task.
The Problem of Missing Metadata
Standard fairness interventions, such as demographic parity or equalized odds, require calculating statistics across different protected groups. If you don't know who belongs to which group, you cannot calculate the disparity, let alone minimize it.
As Joshua Lee, Yuheng Bu, and their colleagues point out in A Maximal Correlation Approach to Imposing Fairness in Machine Learning, this is particularly problematic in transfer learning. A model might be pre-trained on a rich dataset and then adapted to a target domain where sensitive information is strictly restricted. Without a way to enforce fairness during this adaptation, the model may inadvertently carry over or amplify biases from its training phase.
Learning to "Scrub" Representations
The most common strategy for fairness transfer involves learning what "sensitive information" looks like in a source domain and then removing that signature from the target domain's data representations.
In Fair Text Classification with Wasserstein Independence, Thibaud Leteno and his team demonstrate this using an auxiliary sensitive-attribute predictor. They train a model to recognize sensitive traits on a labeled source dataset. They then use this "bias detector" to guide the training of a text classifier on an unlabeled target dataset. By minimizing the Wasserstein distance between the representations of different groups (as identified by the detector), they effectively scrub the sensitive information from the model’s internal logic without ever seeing a sensitive label in the target data.
Causal Disentanglement and Structural Alignment
For more complex data structures, like social networks or citation webs, simple representation scrubbing isn't enough. The bias is often baked into the connections between nodes.
Hourun Li and Yifan Wang address this in their framework, CELL: A Causal Perspective for Fairness-aware Graph Adaptation. Their approach uses causal disentanglement to separate the features of a graph into two distinct parts:
- Fair components: Information that is relevant to the task but independent of sensitive attributes.
- Sensitive components: Information that correlates with protected groups.
By aligning the structure of a labeled source graph with an unlabeled target graph, they can "pseudo-label" the target nodes. This allows the model to learn which parts of the target graph's structure are likely driven by bias and should be ignored, ensuring that the final predictions are based only on the "fair" causal factors.
Detecting Chronic Inequity
Fairness transfer isn't just about building better models; it's also about auditing existing systems where data is sparse. In the context of public utilities, Wei Song and Xiangrui Kong explored how to detect service inequities in electricity restoration.
In Assessing Electricity Service Unfairness with Transfer Counterfactual Learning, they show that patterns of bias learned from "data-rich" scenarios (like massive, high-impact storms where disparities are obvious) can be transferred to "data-sparse" daily operations. This allows researchers to identify subtle, chronic systemic biases that would otherwise be statistically invisible in small, localized datasets.
Why It Works: The Shared Latent Space
The underlying reason these methods work is the assumption of a shared latent space. Even if the specific data points change between domains, the mechanism of bias—how race or gender correlates with certain proxy features—often remains consistent.
By using techniques like Maximal Correlation (HGR), researchers can measure and minimize these dependencies. As Joshua Lee argues, using these correlation-based regularizers provides a smooth, controllable way to trade off accuracy for fairness. It allows the model to find a "middle ground" where it retains as much predictive power as possible while ensuring that its internal representations are independent of the sensitive attributes it learned to recognize in the source domain.