The most intuitive approach to building a fair machine learning model is also the most flawed. For years, practitioners relied on fairness through unawareness (or fairness through blindness)—the practice of simply removing sensitive attributes, such as race, gender, or age, from the training data. The logic seemed airtight: if a model cannot see a protected characteristic, it cannot discriminate based on it.
However, a large body of research has demonstrated that this strategy consistently fails. Real-world data is highly interconnected and filled with redundant encodings. When a sensitive attribute is removed, its predictive signal does not vanish; instead, it leaks into the remaining non-sensitive features, which act as proxies.
For example, a geographic variable like a zip code or an educational metric can correlate strongly with race or socioeconomic status. High-capacity models—ranging from gradient boosted trees (Vargo et al.) to deep neural networks—are exceptionally skilled at identifying and combining these subtle, non-linear traces. Consequently, the model mathematically reconstructs the excluded sensitive information, maintaining biased decision-making patterns under the guise of blindness.
To achieve true algorithmic fairness, researchers have shifted from passive feature exclusion to active proxy mitigation.
The Causal Mechanism of Proxy Leakage
To understand why proxies are so persistent, we must look beyond surface-level statistical correlations. Standard joint distributions cannot tell us why two variables move together, making it difficult to separate benign predictive patterns from discriminatory ones.
Kilbertus et al. formalized this challenge by framing discrimination through causal graphs. By mapping how information flows from a sensitive attribute through intermediate variables to the final prediction, the authors distinguished between legitimate pathways (resolved discrimination) and illegitimate ones (proxy discrimination). This causal lens reveals that a feature is a proxy precisely because it is a downstream casualty or correlate of the protected attribute.
[Sensitive Attribute (e.g., Race)] ---> [Proxy Feature (e.g., Zip Code)] ---> [Model Prediction]
When explicit demographic data is entirely missing due to privacy regulations, detecting these causal pathways becomes a "catch-22." To resolve this, Grari et al. introduced a framework called SRCVAE (Fairness without the Sensitive Attribute via Causal Variational Autoencoder). Instead of throwing up their hands in the absence of sensitive labels, the authors used a causal variational autoencoder to mathematically reconstruct a latent proxy of the missing sensitive attribute from the observable features. This reconstructed proxy is then used to actively guide and de-bias the model during training.
Breaking the Correlation via Representation Learning
If models naturally exploit proxy features to maximize accuracy, how do we force them to stop? The core strategy is to learn data representations that are explicitly "blinded" to proxy correlations before the final prediction is made.
Non-Linear Correlation Minimization
Simple linear correlation metrics fail to capture the complex ways deep networks reconstruct sensitive attributes. To address this, Grari et al. proposed minimizing the Hirschfeld-Gebelein-Rényi (HGR) maximal correlation coefficient in Fairness-Aware Neural Rényi Minimization for Continuous Features. The HGR coefficient looks for the worst-case non-linear transformations of both the model's prediction and the sensitive attribute that maximize their correlation. By minimizing this value, the network actively decouples its predictions from continuous sensitive traits (like age or income), neutralizing even highly non-linear proxies.
Geometric and Distributional Alignment
Another powerful approach is to map data into a latent space where sensitive groups become completely indistinguishable, yet target classes remain highly separable.
- Separability Frameworks: Jang et al. developed the FSNS framework (Achieving Fairness through Separability: A Unified Framework for Fair Representation Learning), which optimizes a latent space to ensure that sensitive attributes are structurally non-separable while target labels are maximally spread apart. The authors extended this concept in their Adversarial Fairness Network (FAIAS) (Adversarial Fairness Network), using an adversarial game to filter out sensitive-relevant features during data selection.
- Avoiding Adversarial Instability: While adversarial training is effective at removing proxy information, it is notoriously unstable. To bypass this, Sinigaglia et al. introduced a method that utilizes Characteristic Function Distance (CFD). Instead of relying on a competing discriminator network, CFD mathematically aligns the probability distributions of the data representations across different sensitive groups, matching all statistical moments simultaneously and achieving state-of-the-art fairness without the training volatility.
Fairness Under Strict Privacy Constraints
In the real world, engineering fair representations is often bottlenecked by legal and privacy frameworks like GDPR, which frequently prohibit companies from collecting or viewing sensitive demographic data in the first place.
To navigate this constraint, Quadrianto et al. proposed a unified framework that treats protected characteristics as "privileged information." Under this paradigm, sensitive data is utilized strictly during the training phase to enforce distribution matching and fairness constraints, but it is completely "recycled" and absent during live predictions. When sensitive data is unavailable even during training, Ni et al. demonstrated that knowledge-sharing frameworks can transfer fairness guarantees across related domains without requiring direct access to protected labels.
However, if a system attempts to bypass missing data by using demographic inference tools to guess protected attributes, errors can easily backfire and compound bias. In the context of ranking algorithms, Olulana et al. demonstrated that post-processing re-ranking strategies are significantly more robust to these inference errors than in-processing models that attempt to integrate noisy fairness constraints directly into the model's primary learning loop.
Expanding to Complex Domains and Verification
As machine learning architectures evolve, the proxy problem manifests in more complex ways, requiring specialized solutions.
Graph-Structured Data
In graph neural networks, proxy leakage occurs not just through individual node features, but also through the network's topology due to homophily (the tendency of similar individuals to connect). To counteract this, Luo et al. designed FairGT, a fairness-aware Graph Transformer. FairGT uses a specialized eigenvector selection strategy for structural encoding alongside a sensitive feature complete graph, successfully decoupling sensitive attributes from both the node features and the underlying graph structure.
Formal Verification
Most fairness techniques are empirical, meaning they are validated on a specific test dataset but offer no guarantees for unseen data. To bridge this gap, Guo et al. introduced formal verification techniques to the fairness domain. By using geometric verification algorithms, the authors can mathematically prove whether a neural network is truly invariant to proxy features within defined input boundaries. Crucially, this verification can be performed on a black-box model without requiring access to the original, often proprietary training data.
Ultimately, these diverse methodologies converge on a single conclusion: fairness cannot be achieved by closing our eyes to sensitive data. Because high-capacity models will always exploit the redundant encodings hidden within proxy features, achieving equity requires explicitly modeling, identifying, and mathematically dismantling those hidden pathways.
Key Foundational Work
Avoiding Discrimination through Causal Reasoning (2017) — contributed 1 concept elements to this direction.
Recycling Privileged Learning and Distribution Matching for Fairness (2017) — contributed 1 concept elements to this direction.
Discrimination-Aware Classifiers for Student Performance Prediction (2015) — contributed 1 concept elements to this direction.